What a speculative decoder gives away for free

Three research records, published as they stood at their pinned commits. The central question came back negative at the registered reader, with one label space open. What holds is narrow, exact and checkable to the byte: every number on these pages drawn from the public record carries the file, the line and the digest it came from.

Existence · S1 and S3
ValueCI95
S1 replicationF-15+0.0141[+0.0081, +0.0203]
S3 primaryF-17+0.0147[+0.0113, +0.0182]
S3 replicationF-17+0.0173[+0.0133, +0.0221]
  • top-1 margins over the proxy free baseline at the pinned L1 cells (F-15, F-17), replicated on fresh seeds
  • absorbed by the full free baseline at the registered reader (F-24) and not overturned by three stronger readers (F-28, F-29)
In plain wordsDid the stream tell a probe anything? A little against the partial baseline, on both sources and on fresh seeds. Against the full free baseline, at the reader registered for the decision, no margin above the floor set in advance, and no more than a scrambled block of the same size adds; one grouped label space stays open at both sources.
README.md:164 · c7c4f9562e72
Conversion · S1 only
ValueCI95
arm4-vs-arm1, seed 1F-19+0.006860[-0.015024, +0.025572]
  • emitted tokens per verify call against vanilla, null against the bar set in advance (F-19, F-21)
  • K-3 fired at the first attempt
  • never run at S3
In plain wordsDid that buy more emitted tokens per check? Not measurably: both registered attempts fell short of the bar set in advance, their intervals against the plain speculative decoder include zero, and the second lost clearly to the same head fed only free features.
sideband/FINDINGS.md:214 · d147e8c0cd6b
Reconstruction · S1 only
ValueCI95
delta_freeF-22+0.204984[+0.153698, +0.288162]
  • linear recovery of the anchor hidden state from the stream over 84 clusters (F-22), a qualified pass at the registered gate
  • the dated erratum reads that target as free information (F-23)
In plain wordsCould the target's internal state at the anchor be rebuilt from the stream? Partly, and better than the free baseline; but the dated erratum showed that this state follows from the free next token distribution, so recovering it recovers nothing the decoder lacked.
README.md:492 · c7c4f9562e72

How to read these pages

The three records
sideband asked whether the stream of memory addresses that weight serving hardware sees for free could help a speculative decoder guess better. amber asked whether a draft model's corrections make a byte exact archive of a larger model's text. lockstep built a pure integer coder, which amber carries as an option, and proposed the container such an archive would need.
Speculative decoding
A small draft model proposes several tokens at once; the large target model checks them in one pass. Accepted tokens are emitted, the first rejected one is replaced by a draw from the target's own corrected distribution, so the output is distributed exactly as the target alone would produce. A better guesser buys speed and never changes what can come out.
A finding (F-N)
One numbered entry in a record's findings log, written when the result landed and never rewritten afterwards. A correction is added beside it with a date. In sideband, each finding names the conditions it was measured under and the committed files its numbers come from.
A kill criterion (K-N)
A way the idea could fail, written down before the data existed, and where a size of effect applied, that size fixed before the run it governed. FIRED means the record met that failure; "did not fire" means it did not; UNTESTED means no run ever tested it.
The free baseline
Everything the decoder already knows at the moment it drafts: the target's own next token probabilities, its top candidates and its internal state at the last verified position. A signal in the stream only counts if it adds to that. The first reads used a partial version of this baseline; the deciding read used the full one.
Intervals
Every value drawn with a line through it is a point estimate with a 95% bootstrap interval. The bootstrap resamples whole prompts, not tokens, so the many positions inside one prompt are not counted as independent evidence.
S1 and S3
The two places the stream was read. S1: on a dense Llama model, the slots of the units each tapped layer activated most strongly, which needs the activation values to rank them, so it is not a pure address stream. S3: on a mixture of experts model (OLMoE), which experts each token was routed to, a pure address stream with no values in it.
L1, L2, L3
The three label spaces the probes predict. L1 is sixteen classes, fifteen coarse concept buckets and one glue class for function words; L2 is about five hundred finer concepts plus the glue class; L3 groups the fine buckets into sets that stay coherent across transitions, two of which hold most of the concepts, which is why the record calls L3 a lopsided lens.
Status words
The record's conclusion labels, in capitals: ABSORBED, SURVIVES, NOT-SHOWN, SPLIT and the rest. sideband's conclusion labels were pinned before the runs they govern, with the exceptions the record names; the rest describe what a later entry did. The glossary gives the record's definition; the plain words beside each use on this page are the page's own.
Chips
The small dotted boxes are provenance chips: the file, the line and the first twelve hex characters of that file's sha256 at the tag. Click one to open those lines in the repository at the release tag. The last section of the front page shows how to check a number against the repository.
The pages
The front page holds the verdicts, the record's own opening, the kill board, the charts, the three parts, where the work sits among prior work, and how to check a number. The findings, the errata, the limits, the modules and the glossary each have a page of their own; the bar pinned to the top of every page moves between them. Everything is also on one page, for searching and printing, linked from the foot of every page.

The plain words explanations, the section intros and the reading guide are this page's own wording, not the record's; they add no claim and carry no chip.

The question and the answer, in the record's words

The two paragraphs that open the repository's README at tag v1.0.0, quoted as written: what the three records asked, and what they found. Everything on these pages unfolds from them. The numbers they state are the record's own and carry its chips; the verdicts in the tiles above are this page's wording.

When a large language model generates text, a small draft model can propose several tokens at once and the large model, the verifier, checks them in one pass; the verifier's rejection rule keeps the output exactly its own, so a better draft buys speed and never changes a result. Something also has to fetch the large model's weights, and whatever serves them, a memory device or a tier holding a mixture-of-experts model's experts, sees which weight groups or experts each token touched, as a byproduct, without reading an activation value or a logit. sideband, the central experiment, asked whether that address stream tells the draft model anything the decoder does not already have, and whether it converts into more emitted tokens. amber asked whether the draft model's rank corrections toward target-generated text make a byte-exact archive of that text. lockstep built a pure-integer coder with conformance vectors, which amber carries as an option, and proposed the model-pinned container such an archive would need.

README.md:6–19 · c7c4f9562e72

The answers are negative or narrow. sideband tested two streams, S1, a value-aware tap on a Llama-3.1-8B target, and S3, the pure address stream of an OLMoE model's routers (both described in the next section). Its probes do find a small signal: at the pinned L1 cells, with a linear probe, the stream raises top-1 accuracy at predicting the concept bucket of the token two positions ahead by +0.0141 at S1 (the replication harvest) and +0.0147 / +0.0173 at S3 (primary / replication) over a summary of what the decoder already knows (F-15, F-17). Against the fuller free baseline the record then built (the logged summaries of the target's next-token distribution, its hashed top-32 token ids and its anchor hidden state, as this pipeline encodes them), read by the registered reader mlp256, the stream's own margin at those L1 cells is about +0.003 (F-24's words), under the +0.005 floor set in advance, and no material amount more than a scrambled block of the same size adds; at L2 the same reading holds (F-25, read under a rider amendment the record calls "a post-hoc reroute around a fired kill condition"); three stronger probe readers, run later at both sources, did not overturn the L1 result (F-28 at S3, where they also read the L3 lenses; F-29 at S1), and L3, the grouped label space, stays open at both. At S1 two registered attempts to convert the signal into emitted tokens per verify call were null against the bar set in advance (F-19, F-21), and the hidden state its reconstruction lane targeted is an exact function of the target's own next-token distribution, free information (F-23); conversion and reconstruction were never run at S3. amber's codec decodes its archives byte-exactly under an fp32 contract on one box (200 of 200 on the GPU, 24 of 24 on the CPU), and its errata pass corrected 16 of the 21 claims it checked. lockstep's byte-model coder and its conformance vectors stand; the model-backed profile was never built, and the economics behind it did not survive measurement.

README.md:21–48 · c7c4f9562e72

Seven ways it could die, written before the data

Before the data, seven ways the idea could fail were written down, and where a size of effect applied, that size was fixed before the run it governed. Two fired: K-2, the stream adds no material amount beyond what the decoder already knows at the registered reader, with one grouped label space left open, and K-3, the signal did not convert into more emitted tokens per check. K-1 did not fire; K-7 did not fire as written, while its gate health diagnostic fired and was routed to an open question. K-4, K-5 and K-6 were never tested: no decoder width sweep ran on real telemetry, nothing in the record is a benchmark harness, and the premise's economics were never reached because both conversion attempts were null. Each card shows the criterion's statement in one line, its latest status entry in the record's own words, and the findings it cites.

K-1

Rejections aren't cross-bucket (Phase 0a)

did not fire

If < ~25% of real rejections cross bucket boundaries at every sane k, concept steering has no room regardless of decoder quality. sideband/docs/WHY_THIS_DIES.md:8 · b09c62345965

In plain wordsIf the target rejected the draft mostly on tokens that sit in the same concept bucket as the right answer, steering at the concept level would have nothing to fix. Measured on the real model pair, most rejections cross bucket boundaries, so this did not fire.
Latest entry
2026-08-29 · outcome

OUTCOME (2026-08-29): did not fire.

F-2 F-8 F-10
sideband/docs/WHY_THIS_DIES.md:13 · b09c62345965
K-2

Telemetry adds nothing beyond free information (Phase 3, OQ-1)

FIRED

The load-bearing risk. sideband/docs/WHY_THIS_DIES.md:60 · b09c62345965

In plain wordsThe central risk: the stream might carry nothing the decoder does not already know for free. Against a partial free baseline a small margin showed; against the full free baseline, at the reader registered for the decision, the margin over a scrambled block of the same size stayed under the floor at the pinned cells of both sources, and three stronger readers did not overturn that; one grouped label space stays open. This fired.
Latest entry
2026-10-17 · status

STATUS (2026-10-17, appended; F-29, S1 READER RE-TEST — the stronger-reader residual named under the S1 FIRED status of 2026-09-06 tested at the S1 gate cell under the three F-28 readers R-N (hz-rescaled), R-T (two-tower) and R-L (longer budget); …

The status text, as written

STATUS (2026-10-17, appended; F-29, S1 READER RE-TEST — the stronger-reader residual named under the S1 FIRED status of 2026-09-06 tested at the S1 gate cell under the three F-28 readers R-N (hz-rescaled), R-T (two-tower) and R-L (longer budget); a NEW registration of 2026-10-16, not an amendment of the reader-and-lens block, whose conditional STAGE S1 stays unrun; the word pinned before any cell trained): S1: K-2-NOT-OVERTURNED-AT-S1 (under readers R-N, R-T and R-L at S1; one more family of readers, not K-2's universal): no reader overturned the registered reader's S1 label under the hz-rescaled, two-tower and longer-budget readers: R-N (hz-rescaled) ABSORBED (D +0.0026 [+0.0008, +0.0043] / +0.0022 [+0.000497, +0.0039]); R-T (two-tower) ABSORBED (D +0.000339 [-0.000377, +0.0011] / +0.0011 [+0.000363, +0.0019]); R-L (longer budget) ABSORBED (D +0.000382 [-0.0009, +0.0016] / -0.000044 [-0.0014, +0.0013]); the stronger-reader residual at S1 L1 is narrowed at the ABSORBED readers (R-N (hz-rescaled), R-T (two-tower), R-L (longer budget)) to readers outside these families and is NOT narrowed at the NOT-SHOWN readers (none); the kill of 2026-09-06 stands as a statement about the registered reader, now with three more readers beside it; the S3 word K-2-ISOLATED (final) does not move.

F-6 F-12 F-14 F-15 F-17 F-18 F-22 F-24 F-25 F-28 F-29
sideband/docs/WHY_THIS_DIES.md:198 · b09c62345965
K-3

Probe margins don't convert to acceptance (Phase 4)

FIRED

Information that exists but arrives too noisy, too coarse, or too stale to move 1 - TV(p, q'). sideband/docs/WHY_THIS_DIES.md:239 · b09c62345965

In plain wordsA signal can exist and still be useless: too noisy, too coarse or too late to change what the draft proposes. The first registered attempt to turn the probe margins into more emitted tokens per check fell short of the bar set in advance. This fired.
Latest entry
2026-09-05 · status

STATUS (2026-09-05, appended): K-3 FIRED at Attempt #1 - F-19 (2026-09-03): adaptive block scheduling on the frozen S1 probes posted +0.006860 CI95 [-0.015024, +0.025572] tokens per verify call against vanilla (bar: …

The status text, as written

STATUS (2026-09-05, appended): K-3 FIRED at Attempt #1 - F-19 (2026-09-03): adaptive block scheduling on the frozen S1 probes posted +0.006860 CI95 [-0.015024, +0.025572] tokens per verify call against vanilla (bar: CI lower > 0 AND point >= +0.062), the pairwise budget guard breached twice by mechanism.

F-4 F-19 F-21 F-22
sideband/docs/WHY_THIS_DIES.md:250 · b09c62345965
K-4

The sideband is too narrow for the decoder it needs (Phase 5, F-5)

UNTESTED

Ungated, a sub-80%-accuracy decoder is worse than nothing; gated, it's harmless but also useless. sideband/docs/WHY_THIS_DIES.md:272 · b09c62345965

In plain wordsA decoder reading the stream must be accurate enough to help. Ungated, a weak one makes drafting worse; gated, it does nothing. Never tested: the widths were priced on paper only, and no decoder width sweep ran on real telemetry.
Latest entry
2026-09-05 · status

STATUS (2026-09-05, appended to K-4, K-5, K-6 together): UNTESTED.

F-5
sideband/docs/WHY_THIS_DIES.md:298 · b09c62345965
K-5

Wall-clock eats the win

UNTESTED

Acceptance gains are real but the sidecar + gating latency inside the draft loop, or the KV-crop engineering, costs more than the accepted tokens buy. sideband/docs/WHY_THIS_DIES.md:278 · b09c62345965

In plain wordsEven a real acceptance gain can lose on the clock if reading the stream and gating inside the draft loop cost more than the accepted tokens save. Never tested: nothing in the record is a benchmark harness, and the record makes no timing claim of its own.
Latest entry
2026-09-05 · status

STATUS (2026-09-05, appended to K-4, K-5, K-6 together): UNTESTED.

sideband/docs/WHY_THIS_DIES.md:298 · b09c62345965
K-6

The premise's economics don't exist

UNTESTED

The device this helps (weight-serving memory with a sideband) only makes sense in MoE/offload regimes (D-1 killed dense-resident). sideband/docs/WHY_THIS_DIES.md:290 · b09c62345965

In plain wordsThe hardware this would help is a memory device that serves weights and sees their addresses, which only makes sense where the weights move during generation (mixture of experts or offload settings). Never tested as an economic premise: both conversion attempts were null, so the question was never posed with a positive number in hand; the record's horizon read left it open, with its datum pointing toward the alternative, not the premise.
Latest entry
2026-10-04 · status

STATUS (2026-10-04, appended; the horizon read at the R2 encoding): K-6 stays UNTESTED.

F-26
sideband/docs/WHY_THIS_DIES.md:385 · b09c62345965
K-7

Glue-bucket tax can't be neutralized in real vocab space

did not fire

D-7's exclusion worked because the synthetic function bucket is labeled. sideband/docs/WHY_THIS_DIES.md:433 · b09c62345965

In plain wordsThe synthetic world could set function word buckets aside because they were labelled. If real text had no such separable buckets, the earlier results would need reopening. The trigger did not fire as written: real glue buckets are concentrated and identifiable; its gate health diagnostic fired and was routed to an open question.
Latest entry
2026-08-30 · outcome

OUTCOME (2026-08-30): the REOPEN TRIGGER did not fire — concept-level share 0.565/0.610/0.674 by k token-level, 0.607 mass-level at k=512, every CI lower bound >2x the bar (F-11).

F-3 F-4 F-10 F-11 F-12 F-13 F-14
sideband/docs/WHY_THIS_DIES.md:469 · b09c62345965

The numbers, drawn to scale

Each chart draws numbers the record prints, to scale. A dot is the point estimate and the line through it is the 95% interval from a bootstrap over prompt clusters. A hollow dot on a dashed line is a row the record labels as not clearing its bar. The vertical dashed lines mark zero and the bar or floor that was set in advance. The record has no wall clock or speedup figure of its own, so none is drawn here. The table view under each chart lists every row as printed, with its source line.

Existence: the stream's margin in top-1 accuracy, against the proxy baseline

Probe margins at the pinned L1 cells, the concept two positions ahead, over the proxy free baseline (F-15 at S1, F-17 at S3, primary and replication harvests).

In plain wordsThe first read of K-2. Each row is a probe trained to guess the concept bucket two positions ahead, once with the stream and once without it; the value is how much more often the probe with the stream is right. Against this partial baseline the stream helped a little, on both sources and on fresh seeds.
0.000+0.005+0.010+0.015+0.0200F-15 · S1 replication+0.0141F-17 · S3 primary+0.0147F-17 · S3 replication+0.0173F-15 · S1 original grid+0.01630.000+0.010+0.0200F-15 · S1 replication+0.0141F-17 · S3 primary+0.0147F-17 · S3 replication+0.0173F-15 · S1 original grid+0.0163

Bootstrap CI95 over prompt clusters: the 30 clusters held out of a pinned 150/30/30 split of a harvest of 210 prompts. These margins are against the proxy baseline; the next chart is the read that decided K-2.

Table view
RowValue as printedLabelFindingSource
F-15 · S1 replication+0.0141 [+0.0081, +0.0203]F-15README.md:164 · c7c4f9562e72
F-17 · S3 primary+0.0147 [+0.0113, +0.0182]F-17README.md:164 · c7c4f9562e72
F-17 · S3 replication+0.0173 [+0.0133, +0.0221]F-17README.md:164 · c7c4f9562e72
F-15 · S1 original grid+0.0163F-15sideband/FINDINGS.md:178 · d147e8c0cd6b

Existence, second read: against the full free baseline, with the +0.005 floor

The equal-dims gate D at the registered reader mlp256 and the pinned L1 cells, against the full free baseline (the logged next token summaries, hashed top-32 ids and the anchor hidden state). Every interval's upper bound sits below the floor set in advance: K-2 FIRED (F-24).

In plain wordsThe read that decided K-2. The same question, but the probe without the stream now sees everything the decoder has for free, and both probes get the same number of inputs so that extra dimensions alone cannot fake a gain. Every interval sits below the floor: no material margin over the full free baseline at the registered reader, and no more than a scrambled block of the same size adds.
-0.002-0.0010.000+0.001+0.002+0.003+0.004+0.0050+0.005 materiality floorS1 L1 · mlp256 · primary+0.0002S1 L1 · mlp256 · replication+0.0012S3 L1 · mlp256 · primary+0.0003S3 L1 · mlp256 · replication+0.0001-0.0020.000+0.002+0.0040+0.005 materiality floorS1 L1 · mlp256 · primary+0.0002S1 L1 · mlp256 · replication+0.0012S3 L1 · mlp256 · primary+0.0003S3 L1 · mlp256 · replication+0.0001

ABSORBED is the record's word for this read; beside each gate D the record prints the stream's own margin Mt as about +0.003. The stronger readers that followed (F-28 at S3, F-29 at S1) are in the K-2 ladder below.

README.md:32 · c7c4f9562e72
Table view
RowValue as printedLabelFindingSource
S1 L1 · mlp256 · primary+0.0002 [-0.0012, +0.0015]ABSORBED · OQ-17F-24sideband/FINDINGS.md:275 · d147e8c0cd6b
S1 L1 · mlp256 · replication+0.0012 [-0.0003, +0.0028]ABSORBED · OQ-17F-24sideband/FINDINGS.md:275 · d147e8c0cd6b
S3 L1 · mlp256 · primary+0.0003 [-0.0013, +0.0020]ABSORBED · OQ-17F-24sideband/FINDINGS.md:275 · d147e8c0cd6b
S3 L1 · mlp256 · replication+0.0001 [-0.0019, +0.0021]ABSORBED · OQ-17F-24sideband/FINDINGS.md:275 · d147e8c0cd6b

Conversion: emitted tokens per verify call against the bar

Two registered attempts at S1 to turn the signal into accepted tokens at a matched drafting budget (F-19, F-21), each leg as the record prints it. The bar set in advance: CI lower above 0 and the point at or above +0.062. Never run at S3.

In plain wordsWhether the signal buys anything. Two registered attempts used a reader of the stream, frozen probes in the first and a trained acceptance head in the second, to decide, block by block, when to stop drafting early or draft more, and counted emitted tokens per check against the plain speculative decoder at the same drafting budget. Neither reached the bar; both intervals against the plain decoder include zero, and the second lost clearly to the same head fed only free features.
-0.150-0.100-0.0500.000+0.050+0.1000+0.062 barF-19 · arm4-vs-arm1+0.00686F-19 · arm4-vs-arm3+0.006974F-19 · arm4-vs-arm2+0.048863F-21 · C4-vs-C1+0.002664F-21 · C4-vs-C2-0.050425F-21 · C4-vs-C3-0.081846-0.1000.000+0.1000+0.062 barF-19 · arm4-vs-arm1+0.00686F-19 · arm4-vs-arm3+0.006974F-19 · arm4-vs-arm2+0.048863F-21 · C4-vs-C1+0.002664F-21 · C4-vs-C2-0.050425F-21 · C4-vs-C3-0.081846

Each verdict is read over a fresh population of 210 prompts; the seed noise floor below says what the bar was worth at that size. Hollow marks on dashed intervals never reach the bar within their interval. This is the record's conversion quantity, not a wall clock figure.

README.md:102 · c7c4f9562e72
Table view
RowValue as printedLabelFindingSource
F-19 · arm4-vs-arm1+0.006860 [-0.015024, +0.025572]NOT WIN · seed 1 (the registered void-rerun)F-19sideband/FINDINGS.md:214 · d147e8c0cd6b
F-19 · arm4-vs-arm3+0.006974 [-0.014909, +0.025690]NOT WIN · seed 1 (the registered void-rerun)F-19sideband/FINDINGS.md:214 · d147e8c0cd6b
F-19 · arm4-vs-arm2+0.048863 [-0.007393, +0.104327]NOT WIN · seed 1 (the registered void-rerun)F-19sideband/FINDINGS.md:214 · d147e8c0cd6b
F-21 · C4-vs-C1+0.002664 [-0.044719, +0.046727]NOT WIN · seed 3 (the registered void-rerun)F-21sideband/FINDINGS.md:239 · d147e8c0cd6b
F-21 · C4-vs-C2-0.050425 [-0.109388, +0.005772]NOT WIN · seed 3 (the registered void-rerun)F-21sideband/FINDINGS.md:239 · d147e8c0cd6b
F-21 · C4-vs-C3-0.081846 [-0.132484, -0.035567]NOT WIN · seed 3 (the registered void-rerun)F-21sideband/FINDINGS.md:239 · d147e8c0cd6b

Seed noise floor: the CI95 half width a seed change alone produces

The median CI95 half width of the null delta when only the seed changes under the same policy, by population and prompt count (F-27, NOISE_FLOOR.md). Rows at n = 30 are drawn in the muted tone.

In plain wordsHow much a conversion verdict can move by chance alone. Running the same decoder with a different random seed changes the tokens per check number; this is the half width of the interval of that change at each prompt count. The record's reading: at thirty prompts a win the size of the bar appears by seed noise alone in more than one subset in ten on both populations; at two hundred and ten prompts the half width is smaller than the bar.
0.000+0.020+0.040+0.060+0.080+0.100K3 · n = 300.0982K3 · n = 420.0844K3 · n = 840.0603K3 · n = 1260.0493K3 · n = 2100.0385OQ15 · n = 300.1032OQ15 · n = 420.0887OQ15 · n = 840.0632OQ15 · n = 1260.0520OQ15 · n = 2100.03980.000+0.050+0.100K3 · n = 300.0982K3 · n = 420.0844K3 · n = 840.0603K3 · n = 1260.0493K3 · n = 2100.0385OQ15 · n = 300.1032OQ15 · n = 420.0887OQ15 · n = 840.0632OQ15 · n = 1260.0520OQ15 · n = 2100.0398

FLOOR-MATERIAL on both populations; BAR-RESOLVED at 210 prompts, RULE-LEAKS at 30. The null is read across seeds under the same policy against a single comparator; the null at the same seed is unmeasured.

sideband/docs/NOISE_FLOOR.md:125 · 0d5fda202e18 sideband/FINDINGS.md:304 · d147e8c0cd6b sideband/docs/NOISE_FLOOR.md:126 · 0d5fda202e18 sideband/docs/NOISE_FLOOR.md:127 · 0d5fda202e18
Table view
RowValue as printedLabelSource
K3 · n = 300.0982median CI95 half-width of the null deltasideband/docs/NOISE_FLOOR.md:79 · 0d5fda202e18
K3 · n = 420.0844median CI95 half-width of the null deltasideband/docs/NOISE_FLOOR.md:79 · 0d5fda202e18
K3 · n = 840.0603median CI95 half-width of the null deltasideband/docs/NOISE_FLOOR.md:79 · 0d5fda202e18
K3 · n = 1260.0493median CI95 half-width of the null deltasideband/docs/NOISE_FLOOR.md:80 · 0d5fda202e18
K3 · n = 2100.0385median CI95 half-width of the null deltasideband/docs/NOISE_FLOOR.md:80 · 0d5fda202e18
OQ15 · n = 300.1032median CI95 half-width of the null deltasideband/docs/NOISE_FLOOR.md:79 · 0d5fda202e18
OQ15 · n = 420.0887median CI95 half-width of the null deltasideband/docs/NOISE_FLOOR.md:79 · 0d5fda202e18
OQ15 · n = 840.0632median CI95 half-width of the null deltasideband/docs/NOISE_FLOOR.md:80 · 0d5fda202e18
OQ15 · n = 1260.0520median CI95 half-width of the null deltasideband/docs/NOISE_FLOOR.md:80 · 0d5fda202e18
OQ15 · n = 2100.0398median CI95 half-width of the null deltasideband/docs/NOISE_FLOOR.md:80 · 0d5fda202e18

Seed noise floor: the other readouts

The rest of the noise floor note's readouts, each as printed: how often a point win the size of the bar or one leg of the win rule passes by seed noise alone, and the related spreads.

0.000+0.050+0.100+0.150K3 · n = 30 · f_one (one-sided ra…0.1060OQ15 · n = 30 · f_one (one-sided …0.1320K3 · n = 42 · f_one (one-sided ra…0.0605OQ15 · n = 42 · f_one (one-sided …0.0860K3 · n = 84 · f_one (one-sided ra…0.0090OQ15 · n = 84 · f_one (one-sided …0.0230K3 · n = 126 · f_one (one-sided r…0.0005OQ15 · n = 126 · f_one (one-sided…0.0005K3 · n = 210 · f_one (one-sided r…0.0000OQ15 · n = 210 · f_one (one-sided…0.0000K3 · n = 30 · g_one (single-leg f…0.0235OQ15 · n = 30 · g_one (single-leg…0.0375K3 · n = 42 · g_one (single-leg f…0.0185OQ15 · n = 42 · g_one (single-leg…0.02500.000+0.050+0.100+0.150K3 · n = 30 · f_one…0.1060OQ15 · n = 30 · f_one…0.1320K3 · n = 42 · f_one…0.0605OQ15 · n = 42 · f_one…0.0860K3 · n = 84 · f_one…0.0090OQ15 · n = 84 · f_one…0.0230K3 · n = 126 · f_one…0.0005OQ15 · n = 126 · f_one…0.0005K3 · n = 210 · f_one…0.0000OQ15 · n = 210 · f_one…0.0000K3 · n = 30 · g_one…0.0235OQ15 · n = 30 · g_one…0.0375K3 · n = 42 · g_one…0.0185OQ15 · n = 42 · g_one…0.0250

Values as printed in NOISE_FLOOR.md; the table view carries every row with its source line.

Table view
RowValue as printedLabelSource
K3 · n = 30 · f_one (one-sided rate of a +0.062 point win by s0.1060f_one (one-sided rate of a +0.062 point win by seed noise alone)sideband/docs/NOISE_FLOOR.md:91 · 0d5fda202e18
OQ15 · n = 30 · f_one (one-sided rate of a +0.062 point win by s0.1320f_one (one-sided rate of a +0.062 point win by seed noise alone)sideband/docs/NOISE_FLOOR.md:91 · 0d5fda202e18
K3 · n = 42 · f_one (one-sided rate of a +0.062 point win by s0.0605f_one (one-sided rate of a +0.062 point win by seed noise alone)sideband/docs/NOISE_FLOOR.md:92 · 0d5fda202e18
OQ15 · n = 42 · f_one (one-sided rate of a +0.062 point win by s0.0860f_one (one-sided rate of a +0.062 point win by seed noise alone)sideband/docs/NOISE_FLOOR.md:92 · 0d5fda202e18
K3 · n = 84 · f_one (one-sided rate of a +0.062 point win by s0.0090f_one (one-sided rate of a +0.062 point win by seed noise alone)sideband/docs/NOISE_FLOOR.md:92 · 0d5fda202e18
OQ15 · n = 84 · f_one (one-sided rate of a +0.062 point win by s0.0230f_one (one-sided rate of a +0.062 point win by seed noise alone)sideband/docs/NOISE_FLOOR.md:92 · 0d5fda202e18
K3 · n = 126 · f_one (one-sided rate of a +0.062 point win by s0.0005f_one (one-sided rate of a +0.062 point win by seed noise alone)sideband/docs/NOISE_FLOOR.md:93 · 0d5fda202e18
OQ15 · n = 126 · f_one (one-sided rate of a +0.062 point win by s0.0005f_one (one-sided rate of a +0.062 point win by seed noise alone)sideband/docs/NOISE_FLOOR.md:93 · 0d5fda202e18
K3 · n = 210 · f_one (one-sided rate of a +0.062 point win by s0.0000f_one (one-sided rate of a +0.062 point win by seed noise alone)sideband/docs/NOISE_FLOOR.md:93 · 0d5fda202e18
OQ15 · n = 210 · f_one (one-sided rate of a +0.062 point win by s0.0000f_one (one-sided rate of a +0.062 point win by seed noise alone)sideband/docs/NOISE_FLOOR.md:93 · 0d5fda202e18
K3 · n = 30 · g_one (single-leg false positives of the F-19 ru0.0235g_one (single-leg false positives of the F-19 rule)sideband/docs/NOISE_FLOOR.md:96 · 0d5fda202e18
OQ15 · n = 30 · g_one (single-leg false positives of the F-19 ru0.0375g_one (single-leg false positives of the F-19 rule)sideband/docs/NOISE_FLOOR.md:96 · 0d5fda202e18
K3 · n = 42 · g_one (single-leg false positives of the F-19 ru0.0185g_one (single-leg false positives of the F-19 rule)sideband/docs/NOISE_FLOOR.md:97 · 0d5fda202e18
OQ15 · n = 42 · g_one (single-leg false positives of the F-19 ru0.0250g_one (single-leg false positives of the F-19 rule)sideband/docs/NOISE_FLOOR.md:97 · 0d5fda202e18
K3 · n = 84 · g_one (single-leg false positives of the F-19 ru0.0075g_one (single-leg false positives of the F-19 rule)sideband/docs/NOISE_FLOOR.md:97 · 0d5fda202e18
OQ15 · n = 84 · g_one (single-leg false positives of the F-19 ru0.0190g_one (single-leg false positives of the F-19 rule)sideband/docs/NOISE_FLOOR.md:97 · 0d5fda202e18
K3 · n = 126 · g_one (single-leg false positives of the F-19 ru0.0005g_one (single-leg false positives of the F-19 rule)sideband/docs/NOISE_FLOOR.md:98 · 0d5fda202e18
OQ15 · n = 126 · g_one (single-leg false positives of the F-19 ru0.0005g_one (single-leg false positives of the F-19 rule)sideband/docs/NOISE_FLOOR.md:98 · 0d5fda202e18
K3 · n = 210 · g_one (single-leg false positives of the F-19 ru0.0000g_one (single-leg false positives of the F-19 rule)sideband/docs/NOISE_FLOOR.md:98 · 0d5fda202e18
OQ15 · n = 210 · g_one (single-leg false positives of the F-19 ru0.0000g_one (single-leg false positives of the F-19 rule)sideband/docs/NOISE_FLOOR.md:98 · 0d5fda202e18
K3 · reliability r (per-prompt tokens/call across the0.4232reliability r (per-prompt tokens/call across the two seeds)sideband/docs/NOISE_FLOOR.md:58 · 0d5fda202e18
OQ15 · reliability r (per-prompt tokens/call across the0.4172reliability r (per-prompt tokens/call across the two seeds)sideband/docs/NOISE_FLOOR.md:58 · 0d5fda202e18
K3 · full-set null delta (n = 210, one draw)+0.011597full-set null delta (n = 210, one draw)sideband/docs/NOISE_FLOOR.md:81 · 0d5fda202e18
OQ15 · full-set null delta (n = 210, one draw)+0.017473full-set null delta (n = 210, one draw)sideband/docs/NOISE_FLOOR.md:81 · 0d5fda202e18

The K-2 ladder: every reader, level and source, primary harvest

The equal-dims gate D with its CI95 for each read of K-2: the registered reader mlp256 at L1 (F-24) and at L2 and L3 (F-25), the balanced lens and the three stronger readers R-N, R-T and R-L at S3 (F-28), and the same three at the S1 gate cell (F-29). Hollow marks on dashed intervals carry the word ABSORBED or NOT-SHOWN or SPLIT; solid marks carry a SURVIVES word.

In plain wordsThe full baseline reads of K-2 on one chart, primary harvest; the table under it adds the replication. A row is a reader (a small model that tries to extract the signal) at a label space and a source. Only two rows clear the floor with a SURVIVES word, both at S3 through the coarse, lopsided L3 lens, and only against the scrambled twin, whose own cost may be what inflates them. The record's final word is that the signal is isolated there, not that the kill is overturned.
-0.0020.000+0.002+0.004+0.006+0.008+0.0100+0.005 floorS1 L1 · mlp256 · primary+0.0002S3 L1 · mlp256 · primary+0.0003S1 L2 · mlp256 · primary+0.0011S1 L3 · mlp256 · lopsided (the L3…+0.001S3 L2 · mlp256 · primary+0.0018S3 L3 · mlp256 · lopsided (the L3…+0.0041S3 L3 · R-0 · balanced (L3_olmoe_…+0.0019S3 L1 · R-N · primary+0.003S3 L1 · R-T · primary+0.0006S3 L3 · R-N · lopsided (L3_olmoe_…+0.008S3 L3 · R-T · lopsided (L3_olmoe_…+0.0012S3 L3 · R-N · balanced (L3_olmoe_…+0.004S3 L3 · R-T · balanced (L3_olmoe_…+0.0015S3 L1 · R-L · primary+0.0012S3 L3 · R-L · lopsided (L3_olmoe_…+0.005107S3 L3 · R-L · balanced (L3_olmoe_…+0.0015S1 L1 · R-N · primary+0.0026S1 L1 · R-T · primary+0.000339S1 L1 · R-L · primary+0.0003820.000+0.005+0.0100+0.005 floorS1 L1 · mlp256 · primary+0.0002S3 L1 · mlp256 · primary+0.0003S1 L2 · mlp256 · primary+0.0011S1 L3 · mlp256 · lopsided (the L3 group lens) · p…+0.001S3 L2 · mlp256 · primary+0.0018S3 L3 · mlp256 · lopsided (the L3 group lens) · p…+0.0041S3 L3 · R-0 · balanced (L3_olmoe_balanced_v1) · p…+0.0019S3 L1 · R-N · primary+0.003S3 L1 · R-T · primary+0.0006S3 L3 · R-N · lopsided (L3_olmoe_groups_v1) · pri…+0.008S3 L3 · R-T · lopsided (L3_olmoe_groups_v1) · pri…+0.0012S3 L3 · R-N · balanced (L3_olmoe_balanced_v1) · p…+0.004S3 L3 · R-T · balanced (L3_olmoe_balanced_v1) · p…+0.0015S3 L1 · R-L · primary+0.0012S3 L3 · R-L · lopsided (L3_olmoe_groups_v1) · pri…+0.005107S3 L3 · R-L · balanced (L3_olmoe_balanced_v1) · p…+0.0015S1 L1 · R-N · primary+0.0026S1 L1 · R-T · primary+0.000339S1 L1 · R-L · primary+0.000382

The record's final words: K-2-ISOLATED (F-28 STAGE READERS-LONG (S3, final)); K-2-NOT-OVERTURNED-AT-S1 (F-29 (S1)). The replication harvest's rows and the READERS stage's duplicates with fewer digits are in the table view of the next figure.

sideband/FINDINGS.md:656 · d147e8c0cd6b sideband/FINDINGS.md:833 · d147e8c0cd6b
Table view
RowValue as printedLabelFindingSource
S1 L1 · mlp256 · primary+0.0002 [-0.0012, +0.0015]ABSORBED · OQ-17F-24sideband/FINDINGS.md:275 · d147e8c0cd6b
S3 L1 · mlp256 · primary+0.0003 [-0.0013, +0.0020]ABSORBED · OQ-17F-24sideband/FINDINGS.md:275 · d147e8c0cd6b
S1 L2 · mlp256 · primary+0.0011 [-0.0001, +0.0023]ABSORBED · OQ-18F-25sideband/FINDINGS.md:286 · d147e8c0cd6b
S1 L3 · mlp256 · lopsided (the L3 group lens) · primary+0.0010 [-0.0006, +0.0026]NOT-SHOWN · OQ-18F-25sideband/FINDINGS.md:286 · d147e8c0cd6b
S3 L2 · mlp256 · primary+0.0018 [+0.0008, +0.0030]ABSORBED · OQ-18F-25sideband/FINDINGS.md:286 · d147e8c0cd6b
S3 L3 · mlp256 · lopsided (the L3 group lens) · primary+0.0041 [+0.0026, +0.0055]SPLIT · OQ-18F-25sideband/FINDINGS.md:286 · d147e8c0cd6b
S3 L3 · R-0 · balanced (L3_olmoe_balanced_v1) · primary+0.0019 [+0.0006, +0.0032]ABSORBED · STAGE LENSF-28sideband/FINDINGS.md:335 · d147e8c0cd6b
S3 L1 · R-N · primary+0.0030 [+0.0016, +0.004548]ABSORBED · STAGE READERS-LONGF-28sideband/FINDINGS.md:668 · d147e8c0cd6b
S3 L1 · R-T · primary+0.0006 [-0.000207, +0.0014]ABSORBED · STAGE READERS-LONGF-28sideband/FINDINGS.md:669 · d147e8c0cd6b
S3 L3 · R-N · lopsided (L3_olmoe_groups_v1) · primary+0.0080 [+0.0062, +0.0097]SURVIVES-STRONGER · STAGE READERSF-28sideband/FINDINGS.md:490 · d147e8c0cd6b
S3 L3 · R-T · lopsided (L3_olmoe_groups_v1) · primary+0.0012 [-0.000341, +0.0026]ABSORBED · STAGE READERS-LONGF-28sideband/FINDINGS.md:674 · d147e8c0cd6b
S3 L3 · R-N · balanced (L3_olmoe_balanced_v1) · primary+0.0040 [+0.0025, +0.005385]NOT-SHOWN · STAGE READERS-LONGF-28sideband/FINDINGS.md:679 · d147e8c0cd6b
S3 L3 · R-T · balanced (L3_olmoe_balanced_v1) · primary+0.0015 [+0.000334, +0.0027]ABSORBED · STAGE READERS-LONGF-28sideband/FINDINGS.md:681 · d147e8c0cd6b
S3 L1 · R-L · primary+0.0012 [+0.000250, +0.0022]ABSORBED · STAGE READERS-LONGF-28sideband/FINDINGS.md:670 · d147e8c0cd6b
S3 L3 · R-L · lopsided (L3_olmoe_groups_v1) · primary+0.005107 [+0.0034, +0.0069]SURVIVES-STRONGER · STAGE READERS-LONGF-28sideband/FINDINGS.md:678 · d147e8c0cd6b
S3 L3 · R-L · balanced (L3_olmoe_balanced_v1) · primary+0.0015 [+0.000220, +0.0029]ABSORBED · STAGE READERS-LONGF-28sideband/FINDINGS.md:682 · d147e8c0cd6b
S1 L1 · R-N · primary+0.0026 [+0.0008, +0.0043]ABSORBED · S1 READER RE-TESTF-29sideband/FINDINGS.md:833 · d147e8c0cd6b
S1 L1 · R-T · primary+0.000339 [-0.000377, +0.0011]ABSORBED · S1 READER RE-TESTF-29sideband/FINDINGS.md:833 · d147e8c0cd6b
S1 L1 · R-L · primary+0.000382 [-0.0009, +0.0016]ABSORBED · S1 READER RE-TESTF-29sideband/FINDINGS.md:833 · d147e8c0cd6b

The full ladder, every harvest and stage

The same ladder as a table: 50 rows as the record prints them, with the replication harvest included and both digit forms where a stage reprinted a cell.

Table view
FindingStageSourceLevelReaderLensHarvestGate D as printedFloorWordSource line
F-24OQ-17S1L1mlp256primary+0.0002 [-0.0012, +0.0015]+0.005sideband/FINDINGS.md:271 · d147e8c0cd6bABSORBEDsideband/FINDINGS.md:275 · d147e8c0cd6b
F-24OQ-17S1L1mlp256replication+0.0012 [-0.0003, +0.0028]+0.005sideband/FINDINGS.md:271 · d147e8c0cd6bABSORBEDsideband/FINDINGS.md:275 · d147e8c0cd6b
F-24OQ-17S3L1mlp256primary+0.0003 [-0.0013, +0.0020]+0.005sideband/FINDINGS.md:271 · d147e8c0cd6bABSORBEDsideband/FINDINGS.md:275 · d147e8c0cd6b
F-24OQ-17S3L1mlp256replication+0.0001 [-0.0019, +0.0021]+0.005sideband/FINDINGS.md:271 · d147e8c0cd6bABSORBEDsideband/FINDINGS.md:275 · d147e8c0cd6b
F-25OQ-18S1L2mlp256primary+0.0011 [-0.0001, +0.0023]+0.005sideband/FINDINGS.md:286 · d147e8c0cd6bABSORBEDsideband/FINDINGS.md:286 · d147e8c0cd6b
F-25OQ-18S1L2mlp256replication+0.0006 [-0.0007, +0.0020]+0.005sideband/FINDINGS.md:286 · d147e8c0cd6bABSORBEDsideband/FINDINGS.md:286 · d147e8c0cd6b
F-25OQ-18S1L3mlp256lopsided (the L3 group lens)primary+0.0010 [-0.0006, +0.0026]+0.005sideband/FINDINGS.md:286 · d147e8c0cd6bNOT-SHOWNsideband/FINDINGS.md:286 · d147e8c0cd6b
F-25OQ-18S1L3mlp256lopsided (the L3 group lens)replication+0.0012 [-0.0015, +0.0040]+0.005sideband/FINDINGS.md:286 · d147e8c0cd6bNOT-SHOWNsideband/FINDINGS.md:286 · d147e8c0cd6b
F-25OQ-18S3L2mlp256primary+0.0018 [+0.0008, +0.0030]+0.005sideband/FINDINGS.md:286 · d147e8c0cd6bABSORBEDsideband/FINDINGS.md:286 · d147e8c0cd6b
F-25OQ-18S3L2mlp256replication+0.0003 [-0.0009, +0.0013]+0.005sideband/FINDINGS.md:286 · d147e8c0cd6bABSORBEDsideband/FINDINGS.md:286 · d147e8c0cd6b
F-25OQ-18S3L3mlp256lopsided (the L3 group lens)primary+0.0041 [+0.0026, +0.0055]+0.005sideband/FINDINGS.md:286 · d147e8c0cd6bSPLITsideband/FINDINGS.md:286 · d147e8c0cd6b
F-25OQ-18S3L3mlp256lopsided (the L3 group lens)replication+0.0053 [+0.0019, +0.0091]+0.005sideband/FINDINGS.md:286 · d147e8c0cd6bSPLITsideband/FINDINGS.md:286 · d147e8c0cd6b
F-28STAGE LENSS3L3R-0balanced (L3_olmoe_balanced_v1)primary+0.0019 [+0.0006, +0.0032]+0.005sideband/FINDINGS.md:342 · d147e8c0cd6bABSORBEDsideband/FINDINGS.md:335 · d147e8c0cd6b
F-28STAGE LENSS3L3R-0balanced (L3_olmoe_balanced_v1)replication+0.0004 [-0.0011, +0.0018]+0.005sideband/FINDINGS.md:342 · d147e8c0cd6bABSORBEDsideband/FINDINGS.md:336 · d147e8c0cd6b
F-28STAGE READERSS3L1R-Nprimary+0.0030 [+0.0016, +0.0045]+0.005sideband/FINDINGS.md:510 · d147e8c0cd6bABSORBEDsideband/FINDINGS.md:484 · d147e8c0cd6b
F-28STAGE READERSS3L1R-Nreplication+0.0032 [+0.0016, +0.004951]+0.005sideband/FINDINGS.md:510 · d147e8c0cd6bABSORBEDsideband/FINDINGS.md:485 · d147e8c0cd6b
F-28STAGE READERSS3L1R-Tprimary+0.0006 [-0.0002, +0.0014]+0.005sideband/FINDINGS.md:510 · d147e8c0cd6bABSORBEDsideband/FINDINGS.md:486 · d147e8c0cd6b
F-28STAGE READERSS3L1R-Treplication+0.0007 [-0.0002, +0.0015]+0.005sideband/FINDINGS.md:510 · d147e8c0cd6bABSORBEDsideband/FINDINGS.md:486 · d147e8c0cd6b
F-28STAGE READERSS3L3R-Nlopsided (L3_olmoe_groups_v1)primary+0.0080 [+0.0062, +0.0097]+0.005sideband/FINDINGS.md:510 · d147e8c0cd6bSURVIVES-STRONGERsideband/FINDINGS.md:490 · d147e8c0cd6b
F-28STAGE READERSS3L3R-Nlopsided (L3_olmoe_groups_v1)replication+0.0130 [+0.0095, +0.0169]+0.005sideband/FINDINGS.md:510 · d147e8c0cd6bSURVIVES-STRONGERsideband/FINDINGS.md:490 · d147e8c0cd6b
F-28STAGE READERSS3L3R-Tlopsided (L3_olmoe_groups_v1)primary+0.0012 [-0.0003, +0.0026]+0.005sideband/FINDINGS.md:510 · d147e8c0cd6bABSORBEDsideband/FINDINGS.md:491 · d147e8c0cd6b
F-28STAGE READERSS3L3R-Tlopsided (L3_olmoe_groups_v1)replication+0.0008 [-0.0006, +0.0023]+0.005sideband/FINDINGS.md:510 · d147e8c0cd6bABSORBEDsideband/FINDINGS.md:491 · d147e8c0cd6b
F-28STAGE READERSS3L3R-Nbalanced (L3_olmoe_balanced_v1)primary+0.0040 [+0.0025, +0.0054]+0.005sideband/FINDINGS.md:510 · d147e8c0cd6bNOT-SHOWNsideband/FINDINGS.md:492 · d147e8c0cd6b
F-28STAGE READERSS3L3R-Nbalanced (L3_olmoe_balanced_v1)replication+0.0045 [+0.0027, +0.0061]+0.005sideband/FINDINGS.md:510 · d147e8c0cd6bNOT-SHOWNsideband/FINDINGS.md:493 · d147e8c0cd6b
F-28STAGE READERSS3L3R-Tbalanced (L3_olmoe_balanced_v1)primary+0.0015 [+0.0003, +0.0027]+0.005sideband/FINDINGS.md:510 · d147e8c0cd6bABSORBEDsideband/FINDINGS.md:494 · d147e8c0cd6b
F-28STAGE READERSS3L3R-Tbalanced (L3_olmoe_balanced_v1)replication+0.0013 [-0.0002, +0.0028]+0.005sideband/FINDINGS.md:510 · d147e8c0cd6bABSORBEDsideband/FINDINGS.md:494 · d147e8c0cd6b
F-28STAGE READERS-LONGS3L1R-Nprimary+0.0030 [+0.0016, +0.004548]+0.005sideband/FINDINGS.md:699 · d147e8c0cd6bABSORBEDsideband/FINDINGS.md:668 · d147e8c0cd6b
F-28STAGE READERS-LONGS3L1R-Nreplication+0.0032 [+0.0016, +0.004951]+0.005sideband/FINDINGS.md:699 · d147e8c0cd6bABSORBEDsideband/FINDINGS.md:668 · d147e8c0cd6b
F-28STAGE READERS-LONGS3L1R-Tprimary+0.0006 [-0.000207, +0.0014]+0.005sideband/FINDINGS.md:699 · d147e8c0cd6bABSORBEDsideband/FINDINGS.md:669 · d147e8c0cd6b
F-28STAGE READERS-LONGS3L1R-Treplication+0.0007 [-0.000191, +0.0015]+0.005sideband/FINDINGS.md:699 · d147e8c0cd6bABSORBEDsideband/FINDINGS.md:669 · d147e8c0cd6b
F-28STAGE READERS-LONGS3L1R-Lprimary+0.0012 [+0.000250, +0.0022]+0.005sideband/FINDINGS.md:699 · d147e8c0cd6bABSORBEDsideband/FINDINGS.md:670 · d147e8c0cd6b
F-28STAGE READERS-LONGS3L1R-Lreplication+0.0015 [-0.000044, +0.0032]+0.005sideband/FINDINGS.md:699 · d147e8c0cd6bABSORBEDsideband/FINDINGS.md:671 · d147e8c0cd6b
F-28STAGE READERS-LONGS3L3R-Nlopsided (L3_olmoe_groups_v1)primary+0.0080 [+0.0062, +0.0097]+0.005sideband/FINDINGS.md:699 · d147e8c0cd6bSURVIVES-STRONGERsideband/FINDINGS.md:673 · d147e8c0cd6b
F-28STAGE READERS-LONGS3L3R-Nlopsided (L3_olmoe_groups_v1)replication+0.0130 [+0.0095, +0.0169]+0.005sideband/FINDINGS.md:699 · d147e8c0cd6bSURVIVES-STRONGERsideband/FINDINGS.md:673 · d147e8c0cd6b
F-28STAGE READERS-LONGS3L3R-Tlopsided (L3_olmoe_groups_v1)primary+0.0012 [-0.000341, +0.0026]+0.005sideband/FINDINGS.md:699 · d147e8c0cd6bABSORBEDsideband/FINDINGS.md:674 · d147e8c0cd6b
F-28STAGE READERS-LONGS3L3R-Tlopsided (L3_olmoe_groups_v1)replication+0.0008 [-0.0006, +0.0023]+0.005sideband/FINDINGS.md:699 · d147e8c0cd6bABSORBEDsideband/FINDINGS.md:674 · d147e8c0cd6b
F-28STAGE READERS-LONGS3L3R-Llopsided (L3_olmoe_groups_v1)primary+0.005107 [+0.0034, +0.0069]+0.005sideband/FINDINGS.md:699 · d147e8c0cd6bSURVIVES-STRONGERsideband/FINDINGS.md:678 · d147e8c0cd6b
F-28STAGE READERS-LONGS3L3R-Llopsided (L3_olmoe_groups_v1)replication+0.005410 [+0.0034, +0.0078]+0.005sideband/FINDINGS.md:699 · d147e8c0cd6bSURVIVES-STRONGERsideband/FINDINGS.md:678 · d147e8c0cd6b
F-28STAGE READERS-LONGS3L3R-Nbalanced (L3_olmoe_balanced_v1)primary+0.0040 [+0.0025, +0.005385]+0.005sideband/FINDINGS.md:699 · d147e8c0cd6bNOT-SHOWNsideband/FINDINGS.md:679 · d147e8c0cd6b
F-28STAGE READERS-LONGS3L3R-Nbalanced (L3_olmoe_balanced_v1)replication+0.0045 [+0.0027, +0.0061]+0.005sideband/FINDINGS.md:699 · d147e8c0cd6bNOT-SHOWNsideband/FINDINGS.md:680 · d147e8c0cd6b
F-28STAGE READERS-LONGS3L3R-Tbalanced (L3_olmoe_balanced_v1)primary+0.0015 [+0.000334, +0.0027]+0.005sideband/FINDINGS.md:699 · d147e8c0cd6bABSORBEDsideband/FINDINGS.md:681 · d147e8c0cd6b
F-28STAGE READERS-LONGS3L3R-Tbalanced (L3_olmoe_balanced_v1)replication+0.0013 [-0.000167, +0.0028]+0.005sideband/FINDINGS.md:699 · d147e8c0cd6bABSORBEDsideband/FINDINGS.md:681 · d147e8c0cd6b
F-28STAGE READERS-LONGS3L3R-Lbalanced (L3_olmoe_balanced_v1)primary+0.0015 [+0.000220, +0.0029]+0.005sideband/FINDINGS.md:699 · d147e8c0cd6bABSORBEDsideband/FINDINGS.md:682 · d147e8c0cd6b
F-28STAGE READERS-LONGS3L3R-Lbalanced (L3_olmoe_balanced_v1)replication+0.0005 [-0.0007, +0.0018]+0.005sideband/FINDINGS.md:699 · d147e8c0cd6bABSORBEDsideband/FINDINGS.md:682 · d147e8c0cd6b
F-29S1 READER RE-TESTS1L1R-Nprimary+0.0026 [+0.0008, +0.0043]+0.005sideband/FINDINGS.md:851 · d147e8c0cd6bABSORBEDsideband/FINDINGS.md:833 · d147e8c0cd6b
F-29S1 READER RE-TESTS1L1R-Nreplication+0.0022 [+0.000497, +0.0039]+0.005sideband/FINDINGS.md:851 · d147e8c0cd6bABSORBEDsideband/FINDINGS.md:833 · d147e8c0cd6b
F-29S1 READER RE-TESTS1L1R-Tprimary+0.000339 [-0.000377, +0.0011]+0.005sideband/FINDINGS.md:851 · d147e8c0cd6bABSORBEDsideband/FINDINGS.md:833 · d147e8c0cd6b
F-29S1 READER RE-TESTS1L1R-Treplication+0.0011 [+0.000363, +0.0019]+0.005sideband/FINDINGS.md:851 · d147e8c0cd6bABSORBEDsideband/FINDINGS.md:833 · d147e8c0cd6b
F-29S1 READER RE-TESTS1L1R-Lprimary+0.000382 [-0.0009, +0.0016]+0.005sideband/FINDINGS.md:851 · d147e8c0cd6bABSORBEDsideband/FINDINGS.md:833 · d147e8c0cd6b
F-29S1 READER RE-TESTS1L1R-Lreplication-0.000044 [-0.0014, +0.0013]+0.005sideband/FINDINGS.md:851 · d147e8c0cd6bABSORBEDsideband/FINDINGS.md:833 · d147e8c0cd6b

Reconstruction: the anchor hidden state from the stream

delta_free = MSE(FREE) - MSE(TEL), pooled TEST standardized MSE, leg 2 of the Stage-R gate (F-22 regime and GATE line): how much of the verifier's anchor hidden state a linear map recovers from the stream at S1 over the free baseline (F-22), and what the dated erratum then said about that target (F-23).

In plain wordsA different question: can a linear map rebuild the target's internal state at the anchor from the stream? Partly, and better than the free baseline. The dated erratum then showed that this state is itself a function of the free next token distribution, so recovering it recovers nothing the decoder lacked.
delta_free, as F-22 prints it
ValueCI95
delta_freeF-22+0.204984[+0.153698, +0.288162]
  • qualified pass at the registered gate
  • the registered magnitude is inflated about 25% by two test prompts outside the training support
  • closed by erratum: the target is free information
README.md:492 · c7c4f9562e72
the same value, full precision, as the README's worked example prints it
Point0.204983722587619
CI95 lower0.15369761574819202
CI95 upper0.28816195871219474
Clusters84 clusters
README.md:491 · c7c4f9562e72 README.md:492 · c7c4f9562e72

The README's worked example recomputes this value from the committed state file with the record's own estimator; the command is under "Check a number".

Seed replay: the conversion legs run again at seed 6

The F-19 and F-21 legs replayed at a fresh seed (F-27's replay stage): SPREAD-COVERED (R2); FREE-SPREAD-COVERED (R6).

In plain wordsThe conversion legs run again at a fresh seed, to see whether their verdicts move. Each verdict leg stays inside the other seed's interval; one descriptive row, the free arm against the plain speculative decoder on the second population, moves within cross seed noise. The record's word per leg is in the table view.
-0.100-0.0500.000+0.050+0.1000+0.062 barK3 R1 · the vs-vanilla leg at see…+0.01034K3 R2+0.01034K3 R5 · the mechanism-matched fre…+0.010109K3 R6+0.010109OQ15 R1 · the vs-vanilla leg at s…-0.018029OQ15 R2-0.018029OQ15 R5 · the mechanism-matched f…-0.053623OQ15 R6-0.053623OQ15 R8+0.035594-0.1000.000+0.1000+0.062 barK3 R1 · the vs-vanilla leg at seed 6+0.01034K3 R2+0.01034K3 R5 · the mechanism-matched free leg at seed 6+0.010109K3 R6+0.010109OQ15 R1 · the vs-vanilla leg at seed 6-0.018029OQ15 R2-0.018029OQ15 R5 · the mechanism-matched free leg at seed 6-0.053623OQ15 R6-0.053623OQ15 R8+0.035594

Each row is one readout of the replay as printed in F-27, with its guard word, or its label and reading where it has no guard; the payload forms with full digits are in the table view.

sideband/FINDINGS.md:307 · d147e8c0cd6b README.md:102 · c7c4f9562e72
Table view
RowValue as printedLabelFindingSourcePayload form
K3 R1 · the vs-vanilla leg at seed 6+0.010340 [-0.009288, +0.030521]GUARD-HELDF-19 legs at seed 6sideband/FINDINGS.md:307 · d147e8c0cd6bdelta 0.010339764491312131 sideband/results/real/seed_replay_k3.json:274 · c0abad0de759ci_lower -0.009287789405354263 sideband/results/real/seed_replay_k3.json:271 · c0abad0de759ci_upper 0.03052136034268161 sideband/results/real/seed_replay_k3.json:273 · c0abad0de759
K3 R2+0.010340 [-0.009288, +0.030521]MUTUAL-COVERAGE · REPRODUCESF-19 legs at seed 6sideband/FINDINGS.md:307 · d147e8c0cd6bdifference -0.003479266316745111 sideband/results/real/seed_replay_k3.json:300 · c0abad0de759ci_lower -0.03341640897915552 sideband/results/real/seed_replay_k3.json:296 · c0abad0de759ci_upper 0.022758295181837335 sideband/results/real/seed_replay_k3.json:297 · c0abad0de759
K3 R5 · the mechanism-matched free leg at seed 6+0.010109 [-0.009724, +0.030422]GUARD-FREE-HELDF-19 legs at seed 6sideband/FINDINGS.md:307 · d147e8c0cd6bdelta 0.010108907950192147 sideband/results/real/seed_replay_k3.json:413 · c0abad0de759ci_lower -0.009723979468228793 sideband/results/real/seed_replay_k3.json:410 · c0abad0de759ci_upper 0.030421962636767583 sideband/results/real/seed_replay_k3.json:412 · c0abad0de759
K3 R6+0.010109 [-0.009724, +0.030422]MUTUAL-COVERAGE · REPRODUCESF-19 legs at seed 6sideband/FINDINGS.md:307 · d147e8c0cd6bdifference -0.003135242797748905 sideband/results/real/seed_replay_k3.json:437 · c0abad0de759ci_lower -0.03332677584334195 sideband/results/real/seed_replay_k3.json:433 · c0abad0de759ci_upper 0.023050641907821048 sideband/results/real/seed_replay_k3.json:434 · c0abad0de759
OQ15 R1 · the vs-vanilla leg at seed 6-0.018029 [-0.065168, +0.028495]GUARD-HELDF-21 legs at seed 6sideband/FINDINGS.md:307 · d147e8c0cd6bdelta -0.018029001184240556 sideband/results/real/seed_replay_oq15.json:279 · a1ef30730baaci_lower -0.06516819083341481 sideband/results/real/seed_replay_oq15.json:276 · a1ef30730baaci_upper 0.02849547629376607 sideband/results/real/seed_replay_oq15.json:278 · a1ef30730baa
OQ15 R2-0.018029 [-0.065168, +0.028495]MUTUAL-COVERAGE · REPRODUCESF-21 legs at seed 6sideband/FINDINGS.md:307 · d147e8c0cd6bdifference 0.02069284229807966 sideband/results/real/seed_replay_oq15.json:305 · a1ef30730baaci_lower -0.044829608082113684 sideband/results/real/seed_replay_oq15.json:301 · a1ef30730baaci_upper 0.08827744983581459 sideband/results/real/seed_replay_oq15.json:302 · a1ef30730baa
OQ15 R5 · the mechanism-matched free leg at seed 6-0.053623 [-0.107526, -0.002448]GUARD-FREE-HELDF-21 legs at seed 6sideband/FINDINGS.md:307 · d147e8c0cd6bdelta -0.05362327911756548 sideband/results/real/seed_replay_oq15.json:418 · a1ef30730baaci_lower -0.10752618306304125 sideband/results/real/seed_replay_oq15.json:415 · a1ef30730baaci_upper -0.0024481798517206822 sideband/results/real/seed_replay_oq15.json:417 · a1ef30730baa
OQ15 R6-0.053623 [-0.107526, -0.002448]MUTUAL-COVERAGE · REPRODUCESF-21 legs at seed 6sideband/FINDINGS.md:307 · d147e8c0cd6bdifference -0.02822292625817102 sideband/results/real/seed_replay_oq15.json:442 · a1ef30730baaci_lower -0.09900247354888371 sideband/results/real/seed_replay_oq15.json:438 · a1ef30730baaci_upper 0.04234621299227831 sideband/results/real/seed_replay_oq15.json:439 · a1ef30730baa
OQ15 R8+0.035594 [-0.009909, +0.082906]DELTA-MOVES · WITHIN-CROSS-SEED-NOISEF-21 legs at seed 6sideband/FINDINGS.md:307 · d147e8c0cd6b

Arcs and findings

An arc is a stretch of the work with one question; the arcs group the findings the way the record's index does. A finding card opens to the record's first sentence, the conditions it was measured under (the regime), the evidence files it points at, and the numbers it states. The capitalised words are the record's conclusion labels; the plain words beside them are this page's. The 67 findings: 30 from sideband, 16 from amber, 21 from lockstep.

Synthetic world (F-0 .. F-9)

The first arc built the verifier and a synthetic concept-HMM testbed (tokens grouped into ground-truth concept buckets, with a tunable draft) and used it to fix the shapes that every later experiment inherits.

The full summary

The first arc built the verifier and a synthetic concept-HMM testbed (tokens grouped into ground-truth concept buckets, with a tunable draft) and used it to fix the shapes that every later experiment inherits. The verifier is exact by enumeration and at float64 epsilon under measurement (F-0, F-1), so acceptance rate is the only outcome any steering experiment can move. In this world most rejections are cross-bucket (76.1% to 89.3%, F-2), always-on bucket boosting self-cancels unless gated on disagreement (F-3), a block-held prediction ages out unless concept persistence comfortably exceeds gamma (F-4), a narrow sideband with a weak decoder is harmful ungated and harmless gated (F-5), a centered posterior bias dominates argmax boosting (F-7), and dense per-position hints lose 20:1 in a bit ledger (F-9). The structural result is F-6: steering from free information and from a stale telemetry oracle are identical curves here, so this world can never show a telemetry win and the existence question had to move to real models.

Kill-criterion outcome: none adjudicated. K-2 is provably un-adjudicable in this world (F-6); K-4 received only a shape (F-5).

The full kill criterion outcome

Kill-criterion outcome: none adjudicated. K-2 is provably un-adjudicable in this world (F-6); K-4 received only a shape (F-5). The mechanism rules from F-3, F-7 and the "free steering pays" half of F-9 were later qualified on real models (F-12, F-13).

F-0 F-1 F-2 F-3 F-4 F-5 F-6 F-7 F-9 F-12 F-13
sideband/docs/FINDINGS_INDEX.md:134 · aa88354d13cb sideband/docs/FINDINGS_INDEX.md:136 · aa88354d13cb sideband/docs/FINDINGS_INDEX.md:151 · aa88354d13cb

Real Phase 0 (F-10 .. F-14)

The first GPU results ran a Llama-3.1-8B-Instruct NF4 target with a Llama-3.2-1B-Instruct bf16 draft over 210 committed prompts (F-10).

The full summary

The first GPU results ran a Llama-3.1-8B-Instruct NF4 target with a Llama-3.2-1B-Instruct bf16 draft over 210 committed prompts (F-10). Rejections are overwhelmingly bucket-level (residual mass 0.778 to 0.863 by k, F-10), and after glue buckets are excluded by the pre-registered K-7 rule the majority are still concept-to-concept jumps (0.565 to 0.674 by k, F-11), so the addressable share of failures is real. Free-information bucket-bias steering, however, never beats vanilla and is harmful at the confirmatory setting (-0.0098 [-0.0155, -0.0039], F-12); the diagnosis is a gate that is open about 94% of the time over 510 fine buckets and a bucket persistence of 0.027 that makes any block-held prediction stale by construction. Coarsening the gating space restores the gate's selectivity but every arm still loses (F-13), and a perfect non-causal forecast at coarse k converts to at most +0.0069 acceptance through this actuator (F-14).

Kill-criterion outcome: K-1 (rejections not cross-bucket) did not fire (F-10); K-7 (glue tax cannot be neutralized) did not fire as written (F-11).

The full kill criterion outcome

Kill-criterion outcome: K-1 (rejections not cross-bucket) did not fire (F-10); K-7 (glue tax cannot be neutralized) did not fire as written (F-11). The free-information actuator was killed as built (F-12, F-13, F-14), which set the D-5 free baseline at zero on real models and blocked Phase 4 until a different actuator existed; no K-3 verdict was reached here.

F-10 F-11 F-12 F-13 F-14
sideband/docs/FINDINGS_INDEX.md:156 · aa88354d13cb sideband/docs/FINDINGS_INDEX.md:158 · aa88354d13cb sideband/docs/FINDINGS_INDEX.md:172 · aa88354d13cb

Existence probes (F-15 .. F-18)

Phase 3 asked kill criterion K-2's question directly: do probes on free summaries plus telemetry beat probes on free summaries alone at offsets >= 2?

The full summary

Phase 3 asked kill criterion K-2's question directly: do probes on free summaries plus telemetry beat probes on free summaries alone at offsets >= 2? At S1 (weight-slot occupancy taps on the Llama target) the pinned replication triple re-fired at +0.0141 [+0.0081, +0.0203] and all eight offset-2/offset-3 triples re-fired, while free information dies by offset 3 and offset 4 is not established (F-15). At S3 (OLMoE router expert ids, a pure address stream) the pinned gate re-fired at +0.0173 [+0.0133, +0.0221] with the entire firing set reproducing cell-for-cell (F-17). Two registered post-hoc diagnostics then tested the cheapest software substitute, concept histograms of the already generated prefix: at both sources it predicts about nothing and the telemetry margin passes an equal-dimension alignment-destroyed noise control (F-16, F-18), with one honest qualifier that about a tenth of S3's coarse-group margin is genuinely topic-shaped.

Kill-criterion outcome: K-2 survived at S1 (F-15) and at S3 (F-17), both on pinned fresh-seed replications, against the 657-summary PROXY; S2 (KV pages) was never run.

The full kill criterion outcome

Kill-criterion outcome: K-2 survived at S1 (F-15) and at S3 (F-17), both on pinned fresh-seed replications, against the 657-summary PROXY; S2 (KV pages) was never run. Dated 2026-10-01: at the registered reader of the FULL free baseline K-2 FIRED at the pinned L1 cells of both sources (F-24) and at L2 (F-25), with L3 OPEN; see the full-free re-tests arc below. Every entry in this arc carries the same caveat: existence is not an actuator, and K-3 (conversion) was untouched.

F-15 F-16 F-17 F-18 F-24 F-25
sideband/docs/FINDINGS_INDEX.md:178 · aa88354d13cb sideband/docs/FINDINGS_INDEX.md:180 · aa88354d13cb sideband/docs/FINDINGS_INDEX.md:194 · aa88354d13cb

Conversion attempt 1 (F-19)

K-3 Attempt #1 froze the S1 probes and used them to decide, per block, when to stop drafting early or extend the block, comparing five arms at matched drafting budget on 210 fresh prompts with emitted tokens per verify call as the estimand.

The full summary

K-3 Attempt #1 froze the S1 probes and used them to decide, per block, when to stop drafting early or extend the block, comparing five arms at matched drafting budget on 210 fresh prompts with emitted tokens per verify call as the estimand. The telemetry arm landed at +0.006860 [-0.015024, +0.025572] over vanilla against a registered bar of +0.062, and a val-tuned advantage of +0.060 shrank to +0.007 on fresh prompts (F-19). The pairwise budget guard was breached twice by a real mechanism (the extend channel inflates budget on harder prompts), and an information-free gamma-6 control showed raw tokens per call is buyable without information, which is why matched budget was load-bearing.

Kill-criterion outcome: K-3 fired exactly as registered. Per the dated fold-back count amendment this opened a two-lane fold-back (OQ-15, then OQ-16, hard stop after the second); …

The full kill criterion outcome

Kill-criterion outcome: K-3 fired exactly as registered. Per the dated fold-back count amendment this opened a two-lane fold-back (OQ-15, then OQ-16, hard stop after the second); the per-offset gating rules, the staleness map, and the byproduct logs were the registered salvage.

F-19
sideband/docs/FINDINGS_INDEX.md:202 · aa88354d13cb sideband/docs/FINDINGS_INDEX.md:204 · aa88354d13cb sideband/docs/FINDINGS_INDEX.md:215 · aa88354d13cb

Fold-back #1 (F-20, F-21)

OQ-15 trained ridge heads to predict per-position analytic acceptance from the K-3 byproduct logs. The Stage I existence gate passed mechanically at delta_noise +0.000763 [+0.000308, +0.001215], but a five-seed stale-target control (targets replaced by within-prompt permutations, so no block-aligned information exists for any …

The full summary

OQ-15 trained ridge heads to predict per-position analytic acceptance from the K-3 byproduct logs. The Stage I existence gate passed mechanically at delta_noise +0.000763 [+0.000308, +0.001215], but a five-seed stale-target control (targets replaced by within-prompt permutations, so no block-aligned information exists for any head) produced margins of the same magnitude and spread, and the free-features head beat the telemetry head outright (delta_free -0.00286 [-0.00322, -0.00252]); the pinned label is ARTIFACT-DOMINANT and the registered meaning of the pass is removed (F-20). Stage C then drove live stop/extend decisions from the frozen heads on 210 fresh prompts: the telemetry head tied vanilla (+0.002664 [-0.044719, +0.046727]) and lost CI-clear to the same policy on free features (-0.081846 [-0.132484, -0.035567]), the live mirror of F-20's delta_free (F-21). A seed-2 budget-guard breach was traced to a selection-time gap between arms, fixed by a registered re-freeze, and the seed-3 rerun cleared the guard on every leg.

Kill-criterion outcome: by the pinned partition, Stage I PASSED plus Stage C NOT-WIN equals CONVERSION-MODE NULL, the second conversion null on the lane, which hands off to OQ-16 rather than stopping the lane.

The full kill criterion outcome

Kill-criterion outcome: by the pinned partition, Stage I PASSED plus Stage C NOT-WIN equals CONVERSION-MODE NULL, the second conversion null on the lane, which hands off to OQ-16 rather than stopping the lane. One descriptive observation survives without a registered CI: head_free posted the highest tokens per call on both seeds at a budget below vanilla's.

F-20 F-21
sideband/docs/FINDINGS_INDEX.md:220 · aa88354d13cb sideband/docs/FINDINGS_INDEX.md:222 · aa88354d13cb sideband/docs/FINDINGS_INDEX.md:238 · aa88354d13cb

Fold-back #2 (F-22 with DR-1 .. DR-4; closed by F-23)

OQ-16 asked how much of the anchor hidden state h_a (the vector SD^2-style draft steering would inject) survives linear reconstruction from the S1 stream.

The full summary

OQ-16 asked how much of the anchor hidden state h_a (the vector SD^2-style draft steering would inject) survives linear reconstruction from the S1 stream. Stage R PASSED against its own pinned expectation of a dilution null: registered delta_noise +0.176469 [+0.130414, +0.251518] and delta_free +0.204984 [+0.153698, +0.288162] (F-22). The critic round found two out-of-support test prompts inflating those magnitudes by about 25%; under the registered DR-1 train-support clip the deployable reading is tel_vs_trivial +0.250784 [+0.239316, +0.261616], an R^2 of +0.253 for telemetry. DR-2 showed a free hashed encoding of the top-32 next-token ids (FREETOK) reconstructs h_a slightly better than telemetry (-0.004792 [-0.007192, -0.002600]); DR-3 showed telemetry plus that encoding (UNION) beats FREETOK by +0.065180 [+0.062411, +0.067705] on both prompt populations; DR-4's second token comparator (FREETOK2, intended to be stronger) landed COMPARATOR-TOOTHLESS, CI-clear worse than FREETOK (-0.003085 [-0.004984, -0.001280] against FREETOK), so whether telemetry adds information "beyond p" itself remains unpriced.

Kill-criterion outcome: no kill verdict; K-3 stays fired from F-19. Stage A (SD^2-style draft-side injection with arms vanilla / oracle_h / free_inject = FREETOK / tel_inject = UNION) was frozen pre-GPU at commit 2e44105 and closed the same day by the dated OQ-16 erratum (F-23): …

The full kill criterion outcome

Kill-criterion outcome: no kill verdict; K-3 stays fired from F-19. Stage A (SD^2-style draft-side injection with arms vanilla / oracle_h / free_inject = FREETOK / tel_inject = UNION) was frozen pre-GPU at commit 2e44105 and closed the same day by the dated OQ-16 erratum (F-23): the telemetry arms are struck because their target follows exactly from the verifier's own p; the oracle arm is a free-actuator question outside the lane (own registration required if it ever runs); no partition branch is claimed - the limbo clause forbids re-labeling a never-run stage a null and the partition's VOID precedent applies - mechanism TARGET-FREE. The fold-back hard stop holds; no third mechanism is authorized.

F-19 F-22 F-23
sideband/docs/FINDINGS_INDEX.md:244 · aa88354d13cb sideband/docs/FINDINGS_INDEX.md:246 · aa88354d13cb sideband/docs/FINDINGS_INDEX.md:273 · aa88354d13cb

Full-free re-tests (F-24, F-25; added 2026-10-01)

After F-23 the question moved one level up: not "is the reconstruction target free" but "does the existence margin itself survive a FULL free baseline" - the 657 summaries of p_a plus a hashed multi-hot of the anchor's top-32 next-token ids plus the z-scored anchor state h_a, read by a registered AdamW reader at mlp256 (30 epochs, the epoch selected on val; …

The full summary

After F-23 the question moved one level up: not "is the reconstruction target free" but "does the existence margin itself survive a FULL free baseline" - the 657 summaries of p_a plus a hashed multi-hot of the anchor's top-32 next-token ids plus the z-scored anchor state h_a, read by a registered AdamW reader at mlp256 (30 epochs, the epoch selected on val; one to three passes at F-24's pinned L1 cells, more at F-25's L2 / L3 cells, where the per-cell best epochs are in the payloads), with an equal-dims alignment-destroyed twin pricing dilution. OQ-17 (F-24) ran it at the pinned L1 offset-2 cells of both sources on both harvests: every gate D has its CI95 upper below the +0.005 materiality floor (S1 +0.0002 [-0.0012, +0.0015] / +0.0012 [-0.0003, +0.0028]; S3 +0.0003 [-0.0013, +0.0020] / +0.0001 [-0.0019, +0.0021]), the registered ABSORBED condition (every D row, linear and offset-3 included, with its CI95 upper below +0.005 and no fire on either harvest), while telemetry on top of the full free baseline still adds about +0.003 (Mt) CI-clear on three of four harvests - matched by the corrected twin's own gain (Mn_r +0.0032 [+0.0016, +0.0050] / +0.0020 [+0.0002, +0.0037] at S1; the S3 pair is in the entry), under which Mt stops being evidence of telemetry information. The pinned prediction (S1 SURVIVES-FULL-FREE; S3 SPLIT or NOT-SHOWN) failed on both sources. The S1 critic recompute found the registered shift twin NOT inert; a post-hoc diagnostic, registered before it ran, re-read the gate against a corrected prompt-level-deranged twin: TWIN-ROBUST at both sources, so the null does not depend on the control. OQ-18 (F-25) extended the same test to the coarse labels L2 / L3. Its registered leak rider fired at S1 L3 (raw leg 2 +0.0132 against the 0.01 band); a dated amendment replaced that leg with a twin-referenced dependence-excess leg after the fire - disclosed in the entry's preface as a post-hoc reroute around a fired kill condition, fresh only at the other three harvests, with a positive control confirming the amended leg fires on a literal non-causal block; the registered raw leg is reported at every cell and was above its band at two (S1 L3 +0.0132, S1_r L2 +0.0105), and at S1 L3 the amended pass is not a CI-clear null (dE_br +0.0026, upper +0.0054). Labels: S1 L2 ABSORBED, S1 L3 NOT-SHOWN, S3 L2 ABSORBED, S3 L3 SPLIT (fired on the replication harvest only, D +0.0053 [+0.0019, +0.0091] decomposing into Mt -0.0014 [-0.0041, +0.0016] and the twin's own cost Mn_r -0.0067, so the entry's post-hoc line reads it as inflated by up to 0.0067 and not as half a survival), every L3 label read through the lopsided group lens the registration names.

Kill-criterion outcome: K-2 FIRED at the registered reader at the pinned L1 cells of both sources (the 2026-09-06 STATUS in docs/WHY_THIS_DIES.md: …

The full kill criterion outcome

Kill-criterion outcome: K-2 FIRED at the registered reader at the pinned L1 cells of both sources (the 2026-09-06 STATUS in docs/WHY_THIS_DIES.md: the existence layer is NOT SHOWN anywhere at that reader at L1, and the stronger-reader residual is the one named path to overturn) and at L2 at both sources (the 2026-09-14 STATUS); S1 L3 and S3 L3 are OPEN at the full-free level (NOT-SHOWN / SPLIT). F-25 adds no all-space sentence (not all four (source, space) pairs read ABSORBED). Dated errata inside F-15 / F-17 and dated notes inside F-16 / F-18 carry the consequences; the original wording stays beside them.

F-15 F-16 F-17 F-18 F-23 F-24 F-25
sideband/docs/FINDINGS_INDEX.md:284 · aa88354d13cb sideband/docs/FINDINGS_INDEX.md:286 · aa88354d13cb sideband/docs/FINDINGS_INDEX.md:324 · aa88354d13cb

OQ-9 stage 1, prefetch (F-26 and its free-bigram twin; added 2026-10-01; the stratified read's paragraph added 2026-10-02; the generalising free baseline's on 2026-10-03; the horizon read's on 2026-10-04)

With the thesis closed negative at the registered reader, the lane turned to the prefetch half of OQ-9 on the committed OLMoE router harvests - cells the registration pins as OCCUPIED (address-only history tables, cross-layer routing predictors, and draft-conditioned expert prefetch, SP-MoE; the history and cross-layer ids to verify) - measured here with the record's own free baseline and no model forward.

The full summary

With the thesis closed negative at the registered reader, the lane turned to the prefetch half of OQ-9 on the committed OLMoE router harvests - cells the registration pins as OCCUPIED (address-only history tables, cross-layer routing predictors, and draft-conditioned expert prefetch, SP-MoE; the history and cross-layer ids to verify) - measured here with the record's own free baseline and no model forward. The estimand is a hit rate: of the next forward's eight experts per gate, the share inside a resident set of B of 64, linear ridge readers, 30 test prompts. At B = 16, horizon 1, a recency cache covers 0.5197 / 0.5155, the free one-token train table 0.8111 / 0.8036, the routing history alone 0.8064 / 0.8056, and history plus the token 0.9143 / 0.9108 (F-26). The four readings all read SIGNAL: HISTORY +0.2867 / +0.2901 over the cache, ADDRESS-BEYOND-TOKEN +0.1032 [+0.0926, +0.1146] / +0.1072 [+0.0952, +0.1206] over the one-token table (below the +0.05 floor at B = 32 for j = 1: +0.0405 / +0.0434, while the B = 32 margin at j = 3 / 4 reads +0.0683 / +0.0654 and +0.0698 / +0.0669 with CI lower above the floor), INTRA-FORWARD +0.4861 / +0.4816 (gate l from the earlier gates of the same forward), and the perfect-draft ceiling at horizon 2 +0.1556 / +0.1545; stage label PREFETCH-HEADROOM, the seed replication agreeing on every label. The landing critic's one blocking objection was on interpretation: the previous token is exactly as free as the current one, is not in the one-token baseline, and gate 0 routes on it (an unresolved alternative, that the stored gate-0 slot lags one forward, is recorded in F-26's GATE-0 NOTE). That became the free-bigram twin, registered as a dated amendment with both predictions pinned (the critic's: the margin drops to or below the floor) and run on 2026-10-01: with the previous token in the baseline at equal dims the address block still adds +0.0609 [+0.0514, +0.0712] / +0.0659 [+0.0547, +0.0788] (D_eq) and the address stream over the free bigram +0.0661 / +0.0705 (D_add) - SURVIVES; both pinned modal expectations had put D_eq below material (the critic's clause FAILED, the assistant's modal label MISSED) - while the free previous token accounts for 0.4104 / 0.3858 of the re-measured margin (point ratios) and both readings fall below the floor at B = 32 (the twin ran at horizon 1 only; the horizon 2..4 margins are over the one-token table and were never tested against the bigram twin). The critic's coverage caveat is carried on every text: on val the covered-key D_eq is +0.0454 / +0.0472, below the floor, against +0.12..+0.45 on the primary where a token key is unseen (an unregistered diagnostic, not evidence), so the material part of the surviving margin sits where the free tables are blind; the diagnostic locates the margin without naming its mechanism, and nothing in the entry reads any of this as context information in the address stream or as support for the sideband premise. F-26's post-hoc reading (NOT a registered readout; it changes no label and licenses nothing) records that ADDRESS-BEYOND-TOKEN stays near +0.10 at every horizon 1..4 (+0.1032 to +0.0933 primary, +0.1072 to +0.0912 replication) while the token's own margin over the cache decays from +0.2913 to +0.0427 (+0.2881 to +0.0415), with the entry's own qualifier that "routing history" may be standing in for free token history; the twin tested horizon 1 only. The stratified coverage read (2026-10-02; registered as a dated amendment before any readout ran, both runs from the freeze commit fea2431) split the twin's test bigram positions by train-table coverage and added the forward's stored output-distribution summaries (mass16, mass64, entropy; 81 coordinates) as a free block in both arms. Precondition REPRODUCE EXACT on both harvests: the stratified rows sum to the twin's 12,960 committed state rows and the pooled D_eq equals the twin's. Label FALLBACK-LOCATED / PT-SURVIVES on both, the seed replication agreeing, both gates straddling the +0.05 floor at B = 16, j = 1. Q1 (coverage): on the covered positions (0.9013 / 0.8582 of the test positions) D_eq +0.0463 [+0.0398, +0.0541] / +0.0481 [+0.0411, +0.0564] (NOT-SHOWN, the floor inside the CI); on the positions with at least one unseen key (0.0987 / 0.1418) +0.1940 [+0.1596, +0.2280] / +0.1741 [+0.1390, +0.2154] (SIGNAL). Q2 (p_t), with the summaries in both arms on all positions: D_eq^PT +0.0532 [+0.0455, +0.0619] / +0.0565 [+0.0480, +0.0664] (SIGNAL, the floor inside the CI) and D_add^PT +0.0583 [+0.0506, +0.0669] / +0.0609 [+0.0525, +0.0709] (SIGNAL); PT GAIN on all positions +0.0089 / +0.0121, on the covered ones +0.0011 / +0.0008. So the caveat is indicated on the unseen-key positions, not settled: a statement about per-position materiality, not about where the pooled margin accrues, since by the point identity the covered positions carry +0.0417 / +0.0412 of the pooled +0.0609 / +0.0659 (0.686 / 0.626 of it); the union's margin is not confined to the unseen keys (D_add^PT on the covered positions +0.0512 / +0.0525, SIGNAL, the floor inside the CI); and where the margin sits is budget-dependent (at B = 8 the covered positions read SIGNAL, D_eq +0.0642 / +0.0672; at B = 32 every Q2 reading is CI-positive below the floor, D_eq^PT +0.0184 / +0.0200 and D_add^PT +0.0197 / +0.0213, where Q2 would read PT-ABSORBS on positions whose pooled twin margin was already below it). The twin's verdict with p_t is not settled at this power; no stage-2 sentence follows, and F-26's sentence stands. PT-SURVIVES says nothing about F-24's full-free block ((a) ++ TOK ++ z-scored h_a, stored in these harvests), the named stronger free arm for the p_t question, and the reported-only pt_only arm's lambda sits at the grid's lower edge on both harvests. Reported, not a label: the covered split by the smaller key's train occurrences reads D_eq +0.0658 / +0.0683 at one (SIGNAL), +0.0552 / +0.0503 at two to four (SIGNAL, the floor inside the CI) and +0.0438 / +0.0456 at five or more (NOT-SHOWN, the floor inside the CI; 0.7508 / 0.6571 of the positions), so the covered union's reading is the well-covered keys'; the both-unseen stratum (0.0113 / 0.0141 of the positions) reads +0.5068 / +0.4613 and draws 0.3482 / 0.5379 of its positions from one test prompt (prompt 13). Next rung R2, a free baseline whose encoding handles unseen and sparse keys; F-26's second dated status line carries the full readout. The generalising free baseline, R2 (2026-10-03; registered as a dated amendment before any readout ran, both runs from the freeze commit 2fe5926) replaced the twin's count tables by replay-prior smoothed tables (the model's own routers on the token's normalised input embedding as the Dirichlet prior of each key's leave-one-prompt-out count row, one weight k = 128 selected on val) and put the forward's stored final-norm state, reduced to 1024 principal directions on train, into the free arm as a second question. Label R2-FALLBACK-SURVIVES / H-ABSORBS on both harvests, REPRODUCE EXACT, no straddle on the unseen-key gate or either Q2 reading. Q1: against the generalising free block the address block still adds +0.0732 [+0.0623, +0.0834] / +0.0662 [+0.0559, +0.0773] on the unseen-key positions (SIGNAL) while the encoding recovers +0.1971 [+0.1697, +0.2266] / +0.1822 [+0.1557, +0.2167] of hit rate there; the pooled Q1 reading +0.0493 [+0.0425, +0.0577] / +0.0506 [+0.0433, +0.0594] reads NOT-SHOWN / SIGNAL by the rule with the CI95 straddling the floor on both — the one label-bearing reading on which the harvests differ, said in words; the located margin is NOT shown to be table sparsity at this encoding. Q2: with the stored final-norm state in the free arm the address block adds +0.0111 [+0.0078, +0.0150] / +0.0106 [+0.0068, +0.0153] and the address stream +0.0172 [+0.0139, +0.0210] / +0.0162 [+0.0126, +0.0206], both NOT-SHOWN, CI-positive below the floor — no material margin (at or above +0.05) of the address block over the equal-dims free block, nor of the address stream over the free bigram plus the hidden block, is shown (both readings CI-positive below the floor); on the both-unseen stratum (672 / 422 positions, 0.3482 / 0.5379 of them from one test prompt, prompt 13) both Q2 readings are SIGNAL with the floor inside every CI (+0.0859 [+0.0340, +0.1002] / +0.0721 [+0.0261, +0.0926] and +0.0868 [+0.0424, +0.1001] / +0.0709 [+0.0299, +0.0895]); beside the strike, as pinned: D_add^H on the unseen-key positions +0.0399 [+0.0279, +0.0514] / +0.0286 [+0.0195, +0.0386] (NOT-SHOWN; the primary's CI95 containing the floor); stage 2's address-beyond-free motivation is struck by the pinned condition (PREFETCH-HEADROOM stands; K-6 stays UNTESTED). The free-baseline ladder under F-26 is closed as pre-committed. See F-26's third dated status line and the block's OUTCOME in docs/BLUEPRINT.md.

Kill-criterion outcome: none; F-26 is NON-CRITERION for every existing finding and a K-6 datum only (a count, not bytes, latency or a boundary argument; …

The full kill criterion outcome

Kill-criterion outcome: none; F-26 is NON-CRITERION for every existing finding and a K-6 datum only (a count, not bytes, latency or a boundary argument; K-6 stays UNTESTED by its dated STATUS lines of 2026-09-15, 2026-10-01, 2026-10-02, 2026-10-03 and 2026-10-04, K-5 by the 2026-09-05 one). Prior art occupies every cell as pinned. The steer half of OQ-9 is PARKED (no actuator in the record converts). The stratified covered-vs-fallback read of D_eq on the TEST split, with a free arm from the harvest's stored p_t summaries, ran on 2026-10-02 (FALLBACK-LOCATED / PT-SURVIVES on both harvests, both gates straddling the floor: the caveat indicated, not settled); the next rung is R2, a free baseline whose encoding handles unseen and sparse keys; stage 2 (a real draft partner for OLMoE) is an option under its own registration, not a plan. (2026-10-03: R2 ran — R2-FALLBACK-SURVIVES / H-ABSORBS on both harvests; the unseen-key margin +0.0732 [+0.0623, +0.0834] / +0.0662 [+0.0559, +0.0773] survives the generalising free encoding, NOT shown to be table sparsity at this encoding, the pooled Q1 reading +0.0493 [+0.0425, +0.0577] / +0.0506 [+0.0433, +0.0594] straddles the floor on both, NOT-SHOWN / SIGNAL by the rule; with the stored final-norm state in the free arm no material margin (at or above +0.05) of the address block over the equal-dims free block, nor of the address stream over the free bigram plus the hidden block, is shown (both readings CI-positive below the floor), +0.0111 [+0.0078, +0.0150] / +0.0106 [+0.0068, +0.0153] equal-dims and +0.0172 [+0.0139, +0.0210] / +0.0162 [+0.0126, +0.0206] union, beside the strike, as pinned: D_add^H on the unseen-key positions +0.0399 [+0.0279, +0.0514] / +0.0286 [+0.0195, +0.0386] (NOT-SHOWN; the primary's CI95 containing the floor); stage 2's address-beyond-free motivation struck; the ladder closed)

F-24 F-26
sideband/docs/FINDINGS_INDEX.md:334 · aa88354d13cb sideband/docs/FINDINGS_INDEX.md:336 · aa88354d13cb sideband/docs/FINDINGS_INDEX.md:458 · aa88354d13cb

Seed-noise floor (F-27; added 2026-10-04)

Every conversion verdict's CI is cross-prompt at one seed, and the +0.062 bar with the F-19 WIN rule was priced against those CIs; F-27 prices the seed-to-seed noise of the same estimand on the four committed vanilla gamma-4 verdict runs (two seed pairs on two 210-prompt populations), CPU only, pre-registered with the critic's three rounds before the freeze.

The full summary

Every conversion verdict's CI is cross-prompt at one seed, and the +0.062 bar with the F-19 WIN rule was priced against those CIs; F-27 prices the seed-to-seed noise of the same estimand on the four committed vanilla gamma-4 verdict runs (two seed pairs on two 210-prompt populations), CPU only, pre-registered with the critic's three rounds before the freeze. A same-policy seed change moves pooled tokens/call by a CI95 half-width of 0.0385 / 0.0398 at 210 prompts and 0.0982 / 0.1032 at 30 (K3 / OQ15); one seed's per-prompt tokens/call correlates 0.4232 [0.3073, 0.5193] / 0.4172 [0.2990, 0.5224] with the other seed's (0.2944 / 0.3368 within families); at n = 30 a +0.062 point win appears by seed noise alone in 0.1060 / 0.1320 of subsets and a single comparator leg of the F-19 rule passes in 0.0235 / 0.0375 (the guard cannot bind between fixed-gamma vanilla runs, so the guarded rate is the same number). Labels FLOOR-MATERIAL on both populations, BAR-RESOLVED@210, RULE-LEAKS@30; the methods note is docs/NOISE_FLOOR.md. What it is not: a cross-seed, same-policy, single-comparator null, while the record's gate legs are same-seed and all-comparator; that null is unmeasured, and no verdict, label or kill status moves (NON-CRITERION). Kill-criterion outcome: none.

F-19 F-21 F-27
sideband/docs/FINDINGS_INDEX.md:509 · aa88354d13cb sideband/docs/FINDINGS_INDEX.md:511 · aa88354d13cb

Reader-and-lens re-test, stage LENS (F-28; added 2026-10-05)

The one named path to overturn K-2 is a stronger or differently balanced reader; the registration folds OQ-13 (ii)'s balanced-lens read into that re-test and runs it in stages.

The full summary

The one named path to overturn K-2 is a stronger or differently balanced reader; the registration folds OQ-13 (ii)'s balanced-lens read into that re-test and runs it in stages. STAGE LENS re-reads the S3 L3 cell through ONE size-capped lens (Fiedler-order quantile binning of the registered transition affinity into 16 groups of 31 or 32; persistence 0.1033 against the lopsided lens's 0.6748 and k16's 0.2293, less coherent than k16 against each lens's own baseline) at the registered reader, CPU only, pre-registered with four critic rounds and a code review before the freeze. Full-free: ABSORBED (gate D +0.0019 [+0.0006, +0.0032] / +0.0004 [-0.0011, +0.0018]; OFF3-ABSORBED, OFF4-ABSORBED; Mt +0.0042 [+0.0025, +0.0061] / +0.0021 [+0.0002, +0.0041]; Mn_r +0.0023 [+0.0010, +0.0038] / +0.0017 [-0.0000, +0.0035], the corrected twin point-positive on both harvests, CI-clear on the primary only, where it cost -0.0031 / -0.0067 at the lopsided lens) where F-25 read SPLIT. Proxy: LENS-SURVIVES — P +0.0221 [+0.0187, +0.0258] / +0.0195 [+0.0160, +0.0232], 0.307 / 0.267 of F-17's +0.0721 / +0.0730; growth over the committed L2 margin (+0.0206 / +0.0207) NOT-SETTLED on both, combined L3-GROWTH-NOT-SETTLED: existence under a size cap, size not settled. The 657-summary proxy (a) at linear reads about +0.003 above the floor at this lens while the full free baseline at mlp256 reads about +0.04 and recovers the whole proxy margin (at matched capacity: linear (a) +0.0029..+0.0035 / +0.0025..+0.0030 against fplus +0.0270..+0.0281 / +0.0243..+0.0256; mlp256 (a) +0.0046..+0.0048 / +0.0018..+0.0032 against fplus +0.0411..+0.0431 / +0.0360..+0.0398 (primary / replication, ranges over the three seeds)), so P is essentially acc((c)) minus the floor. What it is not: no K-2 word (that is STAGE READERS'); the lopsided-lens numbers stand as registered; no second lens. Kill-criterion outcome: none from this stage.

F-17 F-25 F-28
sideband/docs/FINDINGS_INDEX.md:550 · aa88354d13cb sideband/docs/FINDINGS_INDEX.md:552 · aa88354d13cb

S1 reader re-test (F-29; added 2026-10-17)

The record's one named path to overturn the S1 kill (F-24) was a stronger reader; F-28 built three and ran them at S3 only, and the image audit's walkthrough named the gap.

The full summary

The record's one named path to overturn the S1 kill (F-24) was a stronger reader; F-28 built three and ran them at S3 only, and the image audit's walkthrough named the gap. Lever 1 of docs/PUBLIC_CUT_PLAN.md section 14.2 [withheld: docs/PUBLIC_CUT_PLAN.md not in this repository; see WITHHELD.md] (the operator, 2026-10-16) re-tests the S1 gate cell (L1_concept_k16, offset 2, mlp256) under the same three readers through the frozen reader-lens instrument's s1 stage, 36 cells per harvest, CPU only, pre-registered with three critic rounds and a code review before the freeze; a thin new instrument issues the S1 word from the frozen labels. R-N (hz-rescaled) ABSORBED; R-T (two-tower) ABSORBED; R-L (longer budget) ABSORBED. The S1 word, pinned before any cell trained: K-2-NOT-OVERTURNED-AT-S1 (under readers R-N, R-T and R-L at S1; one more family of readers, not K-2's universal). Numbers per reader in F-29. What it is not: no S3 sentence (the S3 word K-2-ISOLATED (final) does not move); no reader that removes both of F-24's mechanisms at once; no L2 / L3 at S1. Kill-criterion outcome: K-2 not overturned at S1 under this family of readers; the registered-reader kill stands, and the stronger-reader residual is narrowed at the ABSORBED readers (narrowed, not closed).

F-24 F-28 F-29
sideband/docs/FINDINGS_INDEX.md:614 · aa88354d13cb sideband/docs/FINDINGS_INDEX.md:616 · aa88354d13cb
F-0Verifier is exact; steering is lossless by constructionsidebandindex: stands

The verifier reproduces the target distribution exactly even when the draft is arbitrarily steered, so steering can only change speed, never output quality

closed-form induced emission distribution equals the target exactly (enumeration, atol 1e-12, 200 random pairs + adversarial edge cases); property survives arbitrary bucket-bias steering of q. sideband/FINDINGS.md:72 · d147e8c0cd6b

In plain words
index: stands
the registered verdict is unchanged and no later entry narrows it
Regime
mathematical property of the Leviathan rejection scheme; holds for any p, q on a shared vocab, including arbitrarily steered q sideband/FINDINGS.md:70 · d147e8c0cd6b
Source
  • tests/test_specdec.py
  • src/sideband/specdec.py
sideband/FINDINGS.md:71 · d147e8c0cd6b
Date
2026-08-28 sideband/FINDINGS.md:69 · d147e8c0cd6b
Index headline
enumeration at atol 1e-12 over 200 random pairs plus adversarial edge cases (exact; no CI)
1e-12the target exactly (enumeration, atol 1e-12, 200 random pairs +sideband/FINDINGS.md:72 · d147e8c0cd6b
200target exactly (enumeration, atol 1e-12, 200 random pairs +sideband/FINDINGS.md:72 · d147e8c0cd6b
sideband/FINDINGS.md:68 · d147e8c0cd6b sideband/docs/FINDINGS_INDEX.md:101 · aa88354d13cb
F-1Exactness holds at machine epsilon under measurementsidebandindex: stands

Measured exactness sits at float64 epsilon and empirical sampling agrees within noise

worst closed-form deviation |induced - p| = 2.2e-16 (float64 epsilon) with and without steering; empirical max deviation 3.2e-4 over 500k samples (within sampling noise, sigma ~= 7e-4). sideband/FINDINGS.md:79 · d147e8c0cd6b

In plain words
index: stands
the registered verdict is unchanged and no later entry narrows it
Regime
mathematical property; 2000 random (p, q) pairs, vocab 2–64, sparsity-injected q, steered variants with beta in [-6, 9] sideband/FINDINGS.md:77 · d147e8c0cd6b
Source
  • results/e01_verifier_exactness.json
  • experiments/e01_verifier_exactness.py
sideband/FINDINGS.md:78 · d147e8c0cd6b
Date
2026-08-28 sideband/FINDINGS.md:76 · d147e8c0cd6b
Index headline
worst closed-form deviation 2.2e-16; empirical max deviation 3.2e-4 over 500k samples (sigma about 7e-4; no CI)
2.2e-16worst closed-form deviation |induced - p| = 2.2e-16 (float64 epsilon) withsideband/FINDINGS.md:79 · d147e8c0cd6b
3.2e-4without steering; empirical max deviation 3.2e-4 over 500k samplessideband/FINDINGS.md:79 · d147e8c0cd6b
500kempirical max deviation 3.2e-4 over 500k samples (withinsideband/FINDINGS.md:79 · d147e8c0cd6b
7e-4samples (within sampling noise, sigma ~= 7e-4). The floor the projectsideband/FINDINGS.md:79 · d147e8c0cd6b
sideband/FINDINGS.md:75 · d147e8c0cd6b sideband/docs/FINDINGS_INDEX.md:102 · aa88354d13cb
F-2Most rejections are cross-bucket, and more so as the draft degradessidebandindex: stands

Most rejections are cross-bucket and the share grows as the draft degrades; this bounds the addressable share of failures, not the gain

cross-bucket fraction of rejections = 89.3% (draft_alpha=0.5), 83.1% (0.7), 76.1% (0.9); vanilla acceptance 0.546 / 0.618 / 0.694. sideband/FINDINGS.md:85 · d147e8c0cd6b

In plain words
index: stands
the registered verdict is unchanged and no later entry narrows it
Regime
synthetic concept-HMM v0, default WorldConfig except draft_alpha; state-observed target; buckets = ground-truth blocks; gamma=4; 24k positions x 3 seeds sideband/FINDINGS.md:83 · d147e8c0cd6b
Source
  • results/e02_crossbucket_audit.json
  • experiments/e02_crossbucket_audit.py
sideband/FINDINGS.md:84 · d147e8c0cd6b
Date
2026-08-28 sideband/FINDINGS.md:82 · d147e8c0cd6b
Index headline
cross-bucket fraction 89.3% / 83.1% / 76.1% at draft_alpha 0.5 / 0.7 / 0.9; vanilla acceptance 0.546 / 0.618 / 0.694 (no CI)
89.3%cross-bucket fraction of rejections = 89.3% (draft_alpha=0.5),sideband/FINDINGS.md:85 · d147e8c0cd6b
0.5fraction of rejections = 89.3% (draft_alpha=0.5), 83.1% (0.7), 76.1%sideband/FINDINGS.md:85 · d147e8c0cd6b
83.1%of rejections = 89.3% (draft_alpha=0.5), 83.1% (0.7), 76.1% (0.9);sideband/FINDINGS.md:85 · d147e8c0cd6b
0.7= 89.3% (draft_alpha=0.5), 83.1% (0.7), 76.1% (0.9); vanillasideband/FINDINGS.md:85 · d147e8c0cd6b
76.1%= 89.3% (draft_alpha=0.5), 83.1% (0.7), 76.1% (0.9); vanillasideband/FINDINGS.md:85 · d147e8c0cd6b
0.9(draft_alpha=0.5), 83.1% (0.7), 76.1% (0.9); vanilla acceptancesideband/FINDINGS.md:85 · d147e8c0cd6b
0.546(0.7), 76.1% (0.9); vanilla acceptance 0.546 / 0.618 / 0.694. Insideband/FINDINGS.md:85 · d147e8c0cd6b
0.61876.1% (0.9); vanilla acceptance 0.546 / 0.618 / 0.694. In this regimesideband/FINDINGS.md:85 · d147e8c0cd6b
sideband/FINDINGS.md:81 · d147e8c0cd6b sideband/docs/FINDINGS_INDEX.md:103 · aa88354d13cb
F-3Uniform bucket boosting nearly self-cancels; disagreement gating fixes itsidebandindex: stands

Always-on bucket boosting nearly self-cancels; gating the boost on draft-vs-prediction disagreement keeps the gain and skips the function-word tax

always-on boosting peaks at a trivial +0.010 acceptance (stale oracle 0.6294 vs vanilla 0.6194 at beta=0.5) and collapses to 0.4376 at beta=6 (-0.18): wrong-concept positions gain, already-right positions pay the function-word tax, and the two nearly cancel. sideband/FINDINGS.md:92 · d147e8c0cd6b

In plain words
index: stands
the registered verdict is unchanged and no later entry narrows it
Regime
synthetic concept-HMM v0, default WorldConfig; gamma=4; 16k positions; oracle policies Dated clause 2026-10-16: steering mechanism argmax, run_simulation's parameter default; the script passes no steer_mech, so D-10's posterior default did not apply (numbers unaffected; DECISIONS.md D-10, dated note). sideband/FINDINGS.md:90 · d147e8c0cd6b
Source
  • results/e03_oracle_ceiling.json
  • experiments/e03_oracle_ceiling.py
sideband/FINDINGS.md:91 · d147e8c0cd6b
Date
2026-08-28 sideband/FINDINGS.md:89 · d147e8c0cd6b
Index headline
ungated peak +0.010 (0.6294 vs vanilla 0.6194 at beta=0.5), collapse to 0.4376 at beta=6; gated stale oracle 0.6389 at beta=3 (+0.0195), fresh oracle 0.6653 at beta=4 (+0.046) (no CI)
+0.010always-on boosting peaks at a trivial +0.010 acceptance (stalesideband/FINDINGS.md:92 · d147e8c0cd6b
0.6294a trivial +0.010 acceptance (stale oracle 0.6294 vs vanilla 0.6194 atsideband/FINDINGS.md:92 · d147e8c0cd6b
0.6194acceptance (stale oracle 0.6294 vs vanilla 0.6194 at beta=0.5) andsideband/FINDINGS.md:92 · d147e8c0cd6b
0.5oracle 0.6294 vs vanilla 0.6194 at beta=0.5) and collapses tosideband/FINDINGS.md:92 · d147e8c0cd6b
0.43760.6194 at beta=0.5) and collapses to 0.4376 at beta=6 (-0.18):sideband/FINDINGS.md:92 · d147e8c0cd6b
-0.18and collapses to 0.4376 at beta=6 (-0.18): wrong-conceptsideband/FINDINGS.md:92 · d147e8c0cd6b
0.6389the gains and skips the tax: stale oracle 0.6389 at beta=3 (+0.0195, ~2xsideband/FINDINGS.md:92 · d147e8c0cd6b
+0.0195the tax: stale oracle 0.6389 at beta=3 (+0.0195, ~2x the ungated peak),sideband/FINDINGS.md:92 · d147e8c0cd6b
sideband/FINDINGS.md:88 · d147e8c0cd6b sideband/docs/FINDINGS_INDEX.md:104 · aa88354d13cb
F-4Staleness tax: held-bucket steering pays only when persistence beats gammasidebandindex: stands

A prediction held for a whole block helps only when concept persistence comfortably exceeds gamma; otherwise it is net harmful at every offset

stale-minus-vanilla acceptance at offset 1 -> 6: rho=0.6 (run length 2.5): -0.004 -> -0.026 (net harmful at every offset); rho=0.75: +0.004 -> -0.019; rho=0.90: +0.013 -> -0.004; rho=0.97 (run length 33): +0.018 -> +0.019 (no visible decay). sideband/FINDINGS.md:99 · d147e8c0cd6b

In plain words
index: stands
the registered verdict is unchanged and no later entry narrows it
Regime
synthetic concept-HMM v0, default WorldConfig except rho; disagreement-gated beta=2.5; gamma=6; 14k positions; analytic per-position acceptance (unconditional, D-9) Dated clause 2026-10-16: steering mechanism argmax, run_simulation's parameter default; the script passes no steer_mech, so D-10's posterior default did not apply (numbers unaffected; DECISIONS.md D-10, dated note); … sideband/FINDINGS.md:97 · d147e8c0cd6b
The full regime line

synthetic concept-HMM v0, default WorldConfig except rho; disagreement-gated beta=2.5; gamma=6; 14k positions; analytic per-position acceptance (unconditional, D-9) Dated clause 2026-10-16: steering mechanism argmax, run_simulation's parameter default; the script passes no steer_mech, so D-10's posterior default did not apply (numbers unaffected; DECISIONS.md D-10, dated note); one seed (200) per curve.

Source
  • results/e04_staleness_decay.json
  • experiments/e04_staleness_decay.py
sideband/FINDINGS.md:98 · d147e8c0cd6b
Date
2026-08-28 sideband/FINDINGS.md:96 · d147e8c0cd6b
Index headline
stale-minus-vanilla at offsets 1 -> 6: rho=0.6 -0.004 -> -0.026; rho=0.97 +0.018 -> +0.019 (no CI)
0.6acceptance at offset 1 -> 6: rho=0.6 (run length 2.5):sideband/FINDINGS.md:99 · d147e8c0cd6b
2.5at offset 1 -> 6: rho=0.6 (run length 2.5): -0.004 -> -0.026 (netsideband/FINDINGS.md:99 · d147e8c0cd6b
-0.004at offset 1 -> 6: rho=0.6 (run length 2.5): -0.004 -> -0.026 (net harmfulsideband/FINDINGS.md:99 · d147e8c0cd6b
-0.0261 -> 6: rho=0.6 (run length 2.5): -0.004 -> -0.026 (net harmful at everysideband/FINDINGS.md:99 · d147e8c0cd6b
0.75-0.026 (net harmful at every offset); rho=0.75: +0.004 -> -0.019;sideband/FINDINGS.md:99 · d147e8c0cd6b
+0.004(net harmful at every offset); rho=0.75: +0.004 -> -0.019; rho=0.90:sideband/FINDINGS.md:99 · d147e8c0cd6b
-0.019at every offset); rho=0.75: +0.004 -> -0.019; rho=0.90: +0.013 ->sideband/FINDINGS.md:99 · d147e8c0cd6b
0.90offset); rho=0.75: +0.004 -> -0.019; rho=0.90: +0.013 -> -0.004;sideband/FINDINGS.md:99 · d147e8c0cd6b
sideband/FINDINGS.md:95 · d147e8c0cd6b sideband/docs/FINDINGS_INDEX.md:105 · aa88354d13cb
F-5Sideband width: bad decoders are harmful ungated, harmless gatedsidebandindex: stands

A narrow sideband with an inaccurate decoder is worse than no steering when ungated and harmless with a confidence gate

decoder holdout accuracy vs K slots: 0.16 (K=8), 0.38, 0.63, 0.83, 0.98, 1.00 (K=1024). sideband/FINDINGS.md:107 · d147e8c0cd6b

In plain words
index: stands
the registered verdict is unchanged and no later entry narrows it
Regime
synthetic concept-HMM v0, hardened telemetry (support_gain=0.5, support_noise=0.5, background_noise=0.25, top_n=16), n_slots swept; disagreement-gated beta=2.5; gamma=4; 12k positions; sidecar = softmax regression on slot multi-hot Dated clause 2026-10-16: steering mechanism argmax, run_simulation's parameter default; the script passes no steer_mech, so D-10's posterior default did not apply (numbers unaffected; … sideband/FINDINGS.md:105 · d147e8c0cd6b
The full regime line

synthetic concept-HMM v0, hardened telemetry (support_gain=0.5, support_noise=0.5, background_noise=0.25, top_n=16), n_slots swept; disagreement-gated beta=2.5; gamma=4; 12k positions; sidecar = softmax regression on slot multi-hot Dated clause 2026-10-16: steering mechanism argmax, run_simulation's parameter default; the script passes no steer_mech, so D-10's posterior default did not apply (numbers unaffected; DECISIONS.md D-10, dated note).

Source
  • results/e05_sidecar_recovery.json
  • experiments/e05_sidecar_recovery.py
sideband/FINDINGS.md:106 · d147e8c0cd6b
Date
2026-08-28 sideband/FINDINGS.md:104 · d147e8c0cd6b
Index headline
ungated recovery -152% (K=8) to +95% (K=1024); gated tau=0.5 -1% to +99%; decoder accuracy 0.16 to 1.00; stale-oracle ceiling +0.016 (no CI)
0.16decoder holdout accuracy vs K slots: 0.16 (K=8), 0.38, 0.63,sideband/FINDINGS.md:107 · d147e8c0cd6b
0.38holdout accuracy vs K slots: 0.16 (K=8), 0.38, 0.63, 0.83, 0.98, 1.00sideband/FINDINGS.md:107 · d147e8c0cd6b
0.63accuracy vs K slots: 0.16 (K=8), 0.38, 0.63, 0.83, 0.98, 1.00sideband/FINDINGS.md:107 · d147e8c0cd6b
0.83vs K slots: 0.16 (K=8), 0.38, 0.63, 0.83, 0.98, 1.00 (K=1024).sideband/FINDINGS.md:107 · d147e8c0cd6b
0.98vs K slots: 0.16 (K=8), 0.38, 0.63, 0.83, 0.98, 1.00 (K=1024).sideband/FINDINGS.md:107 · d147e8c0cd6b
1.00slots: 0.16 (K=8), 0.38, 0.63, 0.83, 0.98, 1.00 (K=1024). Recovery ofsideband/FINDINGS.md:107 · d147e8c0cd6b
-152%of the stale-oracle gain, ungated (tau=0): -152%, -88%, -22%, +36%,sideband/FINDINGS.md:107 · d147e8c0cd6b
-88%stale-oracle gain, ungated (tau=0): -152%, -88%, -22%, +36%, +91%, +95%sideband/FINDINGS.md:107 · d147e8c0cd6b
sideband/FINDINGS.md:103 · d147e8c0cd6b sideband/docs/FINDINGS_INDEX.md:106 · aa88354d13cb
F-6This world provably cannot show telemetry beating free informationsidebandindex: stands

Steering from the last verified distribution and from a stale telemetry oracle give identical curves, so no synthetic telemetry win is possible

free_lastp (steer from the target's last verified distribution — information every spec decoder gets for free) and oracle_concept_stale produce IDENTICAL curves (both best 0.6389 at beta=3, equal at every beta and gate). sideband/FINDINGS.md:114 · d147e8c0cd6b

In plain words
index: stands
the registered verdict is unchanged and no later entry narrows it
Regime
synthetic concept-HMM v0 — structural property of the world, not a tuning outcome. Dated clause 2026-10-16: the e03 numbers quoted here ran at the argmax mechanism, run_simulation's parameter default (DECISIONS.md D-10, dated note). sideband/FINDINGS.md:112 · d147e8c0cd6b
Source
  • results/e03_oracle_ceiling.json
  • tests/test_steering_pipeline.py
sideband/FINDINGS.md:113 · d147e8c0cd6b
Date
2026-08-28 sideband/FINDINGS.md:111 · d147e8c0cd6b
Index headline
both best 0.6389 at beta=3, equal at every beta and gate (no CI)
0.6389produce IDENTICAL curves (both best 0.6389 at beta=3, equal atsideband/FINDINGS.md:114 · d147e8c0cd6b
3IDENTICAL curves (both best 0.6389 at beta=3, equal at every betasideband/FINDINGS.md:114 · d147e8c0cd6b
sideband/FINDINGS.md:110 · d147e8c0cd6b sideband/docs/FINDINGS_INDEX.md:107 · aa88354d13cb
F-7The posterior mechanism dominates argmax steering at every widthsidebandindex: stands

A centered posterior-mass bias beats single-bucket argmax boosting for every policy and width; the disagreement gate stays load-bearing

replacing the argmax single-bucket boost with a CENTERED posterior-mass bias (beta*(P(b)-1/C) per concept bucket, glue 0) improves every policy on the default world — stale oracle +0.0165 vs +0.0124, fresh oracle +0.0272 vs +0.0255, free_lastp +0.0152 vs +0.0124 over vanilla 0.6197. sideband/FINDINGS.md:121 · d147e8c0cd6b

In plain words
index: stands
the registered verdict is unchanged and no later entry narrows it
Regime
part A: synthetic concept-HMM v0, default WorldConfig, n=12000, seeds 400-401, beta=2.5, disagreement gate; part B: hardened telemetry (support_gain=0.5, support_noise=0.5, background_noise=0.25, top_n=16), n=8000, seeds 401-402 sideband/FINDINGS.md:119 · d147e8c0cd6b
Source
  • results/e06_posterior_mechanism.json
  • experiments/e06_posterior_mechanism.py
  • tests/test_posterior_mechanism.py
sideband/FINDINGS.md:120 · d147e8c0cd6b
Date
2026-08-28 sideband/FINDINGS.md:118 · d147e8c0cd6b
Index headline
stale oracle +0.0165 vs +0.0124, fresh +0.0272 vs +0.0255, free_lastp +0.0152 vs +0.0124 over vanilla 0.6197; posterior oracle 0.6365 vs argmax 0.6343 (no CI)
-1a CENTERED posterior-mass bias (beta*(P(b)-1/C) per concept bucket,sideband/FINDINGS.md:121 · d147e8c0cd6b
+0.0165policy on the default world — stale oracle +0.0165 vs +0.0124, freshsideband/FINDINGS.md:121 · d147e8c0cd6b
+0.0124the default world — stale oracle +0.0165 vs +0.0124, fresh oracle +0.0272sideband/FINDINGS.md:121 · d147e8c0cd6b
+0.0272oracle +0.0165 vs +0.0124, fresh oracle +0.0272 vs +0.0255, free_lastpsideband/FINDINGS.md:121 · d147e8c0cd6b
+0.0255+0.0165 vs +0.0124, fresh oracle +0.0272 vs +0.0255, free_lastp +0.0152 vssideband/FINDINGS.md:121 · d147e8c0cd6b
+0.0152fresh oracle +0.0272 vs +0.0255, free_lastp +0.0152 vs +0.0124 over vanillasideband/FINDINGS.md:121 · d147e8c0cd6b
+0.0124+0.0272 vs +0.0255, free_lastp +0.0152 vs +0.0124 over vanilla 0.6197. Onsideband/FINDINGS.md:121 · d147e8c0cd6b
0.6197free_lastp +0.0152 vs +0.0124 over vanilla 0.6197. On hardened worlds thesideband/FINDINGS.md:121 · d147e8c0cd6b
sideband/FINDINGS.md:117 · d147e8c0cd6b sideband/docs/FINDINGS_INDEX.md:108 · aa88354d13cb
F-8Sampled and analytic cross-bucket estimators agree; mass is now the defaultsidebandindex: stands

The analytic cross-bucket residual-mass estimator matches the sampled fraction with zero sampling variance and becomes the default estimand

the analytic estimator (mean residual-distribution mass outside the proposed token's bucket) tracks the sampled fraction within sampling noise: 0.8924 vs 0.8933 (alpha=0.5), 0.8358 vs 0.8315 (0.7), 0.7624 vs 0.7611 (0.9). sideband/FINDINGS.md:128 · d147e8c0cd6b

In plain words
index: stands
the registered verdict is unchanged and no later entry narrows it
Regime
synthetic concept-HMM v0, default WorldConfig except draft_alpha; gamma=4; 24k positions x 3 seeds (same runs as F-2, reproduced bit-exact after the v0.2.0 pipeline changes) sideband/FINDINGS.md:126 · d147e8c0cd6b
Source
  • results/e02_crossbucket_audit.json
  • src/sideband/pipeline.py
  • src/sideband/real/hf_specdec.py
sideband/FINDINGS.md:127 · d147e8c0cd6b
Date
2026-08-28 sideband/FINDINGS.md:125 · d147e8c0cd6b
Index headline
0.8924 vs 0.8933 (alpha 0.5), 0.8358 vs 0.8315 (0.7), 0.7624 vs 0.7611 (0.9); e03 vanilla 0.6194399309459379 bit-exact
0.8924the sampled fraction within sampling noise: 0.8924 vs 0.8933 (alpha=0.5),sideband/FINDINGS.md:128 · d147e8c0cd6b
0.8933fraction within sampling noise: 0.8924 vs 0.8933 (alpha=0.5), 0.8358 vssideband/FINDINGS.md:128 · d147e8c0cd6b
0.5sampling noise: 0.8924 vs 0.8933 (alpha=0.5), 0.8358 vs 0.8315sideband/FINDINGS.md:128 · d147e8c0cd6b
0.8358noise: 0.8924 vs 0.8933 (alpha=0.5), 0.8358 vs 0.8315 (0.7), 0.7624sideband/FINDINGS.md:128 · d147e8c0cd6b
0.83150.8924 vs 0.8933 (alpha=0.5), 0.8358 vs 0.8315 (0.7), 0.7624 vs 0.7611sideband/FINDINGS.md:128 · d147e8c0cd6b
0.7vs 0.8933 (alpha=0.5), 0.8358 vs 0.8315 (0.7), 0.7624 vs 0.7611sideband/FINDINGS.md:128 · d147e8c0cd6b
0.76240.8933 (alpha=0.5), 0.8358 vs 0.8315 (0.7), 0.7624 vs 0.7611 (0.9). Samesideband/FINDINGS.md:128 · d147e8c0cd6b
0.76110.8358 vs 0.8315 (0.7), 0.7624 vs 0.7611 (0.9). Same estimand,sideband/FINDINGS.md:128 · d147e8c0cd6b
sideband/FINDINGS.md:124 · d147e8c0cd6b sideband/docs/FINDINGS_INDEX.md:109 · aa88354d13cb
F-9Hint-rate ledger v0: dense per-position hints lose 20:1; free steering pays outrightsidebandindex: stands

Priced in bits, dense per-position hints lose 20:1 while free steering saves bits at zero hint cost

vanilla pays 1.270 correction bits per emitted token (0.350 rejections/token). sideband/FINDINGS.md:135 · d147e8c0cd6b

In plain words
index: stands
the registered verdict is unchanged and no later entry narrows it
Regime
synthetic concept-HMM v0, default WorldConfig; posterior mechanism, disagreement gate, beta=2.5, gamma=4, n=16000, seeds 500-501; ideal-coding ledger, accept/reject indicator stream omitted, H1 hint priced per drafted position (conservative) sideband/FINDINGS.md:133 · d147e8c0cd6b
Source
  • results/e07_hint_rate.json
  • experiments/e07_hint_rate.py
sideband/FINDINGS.md:134 · d147e8c0cd6b
Date
2026-08-28 sideband/FINDINGS.md:132 · d147e8c0cd6b
Index headline
vanilla 1.270 correction bits/token; free steering saves 0.100 bits/token; dense hint net -3.18 bits/token; break-even hint budget 0.162 bits/token (no CI)
20:1ledger v0: dense per-position hints lose 20:1; free steering payssideband/FINDINGS.md:131 · d147e8c0cd6b
1.270- finding: vanilla pays 1.270 correction bits persideband/FINDINGS.md:135 · d147e8c0cd6b
0.3501.270 correction bits per emitted token (0.350 rejections/token).sideband/FINDINGS.md:135 · d147e8c0cd6b
0.100steering (zero hint bits) cuts that by 0.100 bits/token — thesideband/FINDINGS.md:135 · d147e8c0cd6b
3.58per-position fresh-concept hint (log2 12 = 3.58 bits per draftedsideband/FINDINGS.md:135 · d147e8c0cd6b
3.34(log2 12 = 3.58 bits per drafted position = 3.34 bits/token aftersideband/FINDINGS.md:135 · d147e8c0cd6b
0.1623.34 bits/token after re-drafts) buys only 0.162 bits/token ofsideband/FINDINGS.md:135 · d147e8c0cd6b
-3.180.162 bits/token of corrections back: net -3.18 bits/token, a 20:1 losssideband/FINDINGS.md:135 · d147e8c0cd6b
sideband/FINDINGS.md:131 · d147e8c0cd6b sideband/docs/FINDINGS_INDEX.md:110 · aa88354d13cb
F-10Real rejections are overwhelmingly cross-bucket; kill criterion K-1 does not firesidebandindex: stands

Real rejections are mostly bucket-level disagreements; kill criterion K-1 does not fire and the quantization choice does not change the verdict

mean acceptance 0.763 over 210 prompts, inside the ~0.6-0.8 sanity band pre-registered in the port checklist (docs/BLUEPRINT.md); NF4-vs-bf16 prompt-paired acceptance delta -0.002 pooled (+0.001 mean-of-rates), CI95 [-0.021, +0.018] over 20 prompts (pinned recipe in tools/verify_phase0a.py). sideband/FINDINGS.md:143 · d147e8c0cd6b

In plain words
index: stands
the registered verdict is unchanged and no later entry narrows it
Regime
REAL MODELS, first GPU result (Phase 0a) — Llama-3.1-8B-Instruct target in NF4 (double-quant, bf16 compute; --load-4bit) + Llama-3.2-1B-Instruct draft in bf16, one shared tokenizer (enforced); gamma=4, temperature 1.0, chat template applied with pinned date; 210 committed short prompts (data/prompts.jsonl, unique topics); seed 0 with per-prompt RNG streams; KV-cached path — … sideband/FINDINGS.md:141 · d147e8c0cd6b
The full regime line

REAL MODELS, first GPU result (Phase 0a) — Llama-3.1-8B-Instruct target in NF4 (double-quant, bf16 compute; --load-4bit) + Llama-3.2-1B-Instruct draft in bf16, one shared tokenizer (enforced); gamma=4, temperature 1.0, chat template applied with pinned date; 210 committed short prompts (data/prompts.jsonl, unique topics); seed 0 with per-prompt RNG streams; KV-cached path — pre-registered cached-vs-uncached A/B PASSED bidirectionally on all six containments (tools/verify_phase0a.py); buckets = k-means over the target's lm_head rows (PCA-128, seed 0), k in {256, 512, 1024}

Source
  • results/real/audit_k256.summary.json
  • results/real/audit_k512.summary.json
  • results/real/audit_k1024.summary.json
  • results/real/sanity_k512.summary.json
  • results/real/sanity_k512_nokv.summary.json
  • results/real/sanity_k512_bf16.summary.json
  • results/real/phase0a_verification.json
  • tools/verify_phase0a.py
  • src/sideband/real/hf_specdec.py
  • docs/WHY_THIS_DIES.md
sideband/FINDINGS.md:142 · d147e8c0cd6b
Date
2026-08-29 sideband/FINDINGS.md:140 · d147e8c0cd6b
Index headline
cross-bucket residual mass 0.778 [0.769, 0.787] (k=256), 0.812 [0.804, 0.820] (k=512), 0.863 [0.856, 0.870] (k=1024); mean acceptance 0.763; 11,287 rejections
0.763- finding: mean acceptance 0.763 over 210 prompts,sideband/FINDINGS.md:143 · d147e8c0cd6b
0.6-0.80.763 over 210 prompts, inside the ~0.6-0.8 sanity bandsideband/FINDINGS.md:143 · d147e8c0cd6b
-0.002NF4-vs-bf16 prompt-paired acceptance delta -0.002 pooled (+0.001sideband/FINDINGS.md:143 · d147e8c0cd6b
+0.001acceptance delta -0.002 pooled (+0.001 mean-of-rates), CI95sideband/FINDINGS.md:143 · d147e8c0cd6b
-0.021-0.002 pooled (+0.001 mean-of-rates), CI95 [-0.021, +0.018] over 20sideband/FINDINGS.md:143 · d147e8c0cd6b
+0.018(+0.001 mean-of-rates), CI95 [-0.021, +0.018] over 20 promptssideband/FINDINGS.md:143 · d147e8c0cd6b
0.778 [0.769, 0.787]RESIDUAL MASS (the F-8 default estimand) = 0.778 [0.769, 0.787] at k=256, 0.812 [0.804,sideband/FINDINGS.md:143 · d147e8c0cd6b
0.812 [0.804, 0.820]estimand) = 0.778 [0.769, 0.787] at k=256, 0.812 [0.804, 0.820] at k=512, 0.863 [0.856,sideband/FINDINGS.md:143 · d147e8c0cd6b
sideband/FINDINGS.md:139 · d147e8c0cd6b sideband/docs/FINDINGS_INDEX.md:111 · aa88354d13cb
F-11Real glue buckets are concentrated and identifiablesidebandindex: standsPASS

Glue concentrates in 2-4 buckets per granularity; after excluding it the majority of rejections are still concept-to-concept jumps

Heading, as written Real glue buckets are concentrated and identifiable; the concept-level share survives (OQ-4 answered; F-10 reopen trigger does not fire)

glue is CONCENTRATED, exactly as the D-7 analogy requires — 2 buckets at k=256 (of 256), 2 at k=512, 4 at k=1024 cover 80% of top-100 frequency mass (82.3%/80.1%/80.0%; concentration flag fires nowhere); decoded members are textbook function tokens (", the, ., a, and, of, to, 's", markdown markers, pronouns). sideband/FINDINGS.md:150 · d147e8c0cd6b

In plain words
PASS
the registered gate passed
index: stands
the registered verdict is unchanged and no later entry narrows it
Regime
REAL MODELS (Phase 0a evidence re-analyzed + Phase 0b vanilla arm) — same pair/regime as F-10 (NF4 8B target, bf16 1B draft, gamma=4, temp 1.0, 210 committed prompts, seed 0); glue identified by the K-7 pre-registered rule (top-100 tokens by frequency in a 53,441-token TARGET-ONLY sampled harvest on the same prompts, greedy 80% frequency-mass cover, rule fixed before any glue set existed); … sideband/FINDINGS.md:148 · d147e8c0cd6b
The full regime line

REAL MODELS (Phase 0a evidence re-analyzed + Phase 0b vanilla arm) — same pair/regime as F-10 (NF4 8B target, bf16 1B draft, gamma=4, temp 1.0, 210 committed prompts, seed 0); glue identified by the K-7 pre-registered rule (top-100 tokens by frequency in a 53,441-token TARGET-ONLY sampled harvest on the same prompts, greedy 80% frequency-mass cover, rule fixed before any glue set existed); mass-level decomposition from the Phase 0b vanilla re-run at k=512, which reproduced the committed audit trajectories BIT-IDENTICALLY (all 210 per-prompt counts, all 11,287 rejection tuples, all float64 masses — run-to-run CUDA determinism on this box/version stack, so the decomposition applies exactly to the F-10 trajectories)

Source
  • results/real/gen_freq.json
  • results/real/glue_k256.json
  • results/real/glue_k512.json
  • results/real/glue_k1024.json
  • results/real/glue_phase0a_analysis.json
  • results/real/phase0b_vanilla_k512.summary.json
  • tools/analyze_glue.py
  • src/sideband/glue.py
  • docs/WHY_THIS_DIES.md
sideband/FINDINGS.md:149 · d147e8c0cd6b
Date
2026-08-30 sideband/FINDINGS.md:147 · d147e8c0cd6b
Index headline
concept-to-concept 0.565 [0.554, 0.575] (k=256), 0.610 [0.600, 0.621] (k=512), 0.674 [0.664, 0.684] (k=1024); glue tokens are 41% of generated text
80%(of 256), 2 at k=512, 4 at k=1024 cover 80% of top-100 frequencysideband/FINDINGS.md:150 · d147e8c0cd6b
82.3%k=1024 cover 80% of top-100 frequency mass (82.3%/80.1%/80.0%;sideband/FINDINGS.md:150 · d147e8c0cd6b
0.565 [0.554, 0.575]from the proposed token's bucket) = 0.565 [0.554, 0.575] at k=256, 0.610 [0.600,sideband/FINDINGS.md:150 · d147e8c0cd6b
0.610 [0.600, 0.621]bucket) = 0.565 [0.554, 0.575] at k=256, 0.610 [0.600, 0.621] at k=512, 0.674 [0.664,sideband/FINDINGS.md:150 · d147e8c0cd6b
0.674 [0.664, 0.684]at k=256, 0.610 [0.600, 0.621] at k=512, 0.674 [0.664, 0.684] at k=1024sideband/FINDINGS.md:150 · d147e8c0cd6b
0.607 [0.597, 0.617]vanilla arm's per-record decomposition) = 0.607 [0.597, 0.617] — sampled and analyticsideband/FINDINGS.md:150 · d147e8c0cd6b
0.25REOPEN TRIGGER (concept-level CI upper < 0.25, token-level at every ksideband/FINDINGS.md:150 · d147e8c0cd6b
0.15at k=512) DOES NOT FIRE; no CI touches the [0.15, 0.35] caution band.sideband/FINDINGS.md:150 · d147e8c0cd6b
sideband/FINDINGS.md:146 · d147e8c0cd6b sideband/docs/FINDINGS_INDEX.md:112 · aa88354d13cb
F-12Free-information steering baseline is ZERO on real modelssidebandindex: standsZERONEGATIVE

Free-information steering never beats vanilla and the confirmatory arm is harmful; the disagreement gate is open about 94% of the time and consecutive-token bucket persistence is about 0.03

Heading, as written Free-information steering baseline is ZERO on real models; the F-3 gate tax reproduces, diagnosed as gate-granularity failure (Phase 0b)

the D-5 denominator is ZERO — no arm beats vanilla (acceptance 0.7628 pooled): best delta +0.0025 [-0.0028, +0.0079] (soft beta=1), CI through zero; the pre-registered CONFIRMATORY arm (sharp beta=3) is definitively NEGATIVE, -0.0098 [-0.0155, -0.0039]; … sideband/FINDINGS.md:157 · d147e8c0cd6b

The full first sentence

the D-5 denominator is ZERO — no arm beats vanilla (acceptance 0.7628 pooled): best delta +0.0025 [-0.0028, +0.0079] (soft beta=1), CI through zero; the pre-registered CONFIRMATORY arm (sharp beta=3) is definitively NEGATIVE, -0.0098 [-0.0155, -0.0039]; harm is dose-monotone in beta for sharp (+0.0009 / -0.0040 / -0.0098 / -0.0237 [-0.0302, -0.0172]) and null-to-slightly-negative for soft (+0.0025 / +0.0002 / -0.0051 / -0.0033).

In plain words
ZERO
no measurable effect; the measured value is zero within its interval
NEGATIVE
the interval sits entirely below zero
index: stands
the registered verdict is unchanged and no later entry narrows it
Regime
REAL MODELS (Phase 0b) — same pair/regime as F-10/F-11 (NF4 8B + bf16 1B, gamma=4, temp 1.0, 210 prompts, seed 0, KV-cached, chat template pinned); k=512 buckets with glue {59, 403} excluded per F-11; steer=bucket, mech=posterior (D-10), gate=disagree; … sideband/FINDINGS.md:155 · d147e8c0cd6b
The full regime line

REAL MODELS (Phase 0b) — same pair/regime as F-10/F-11 (NF4 8B + bf16 1B, gamma=4, temp 1.0, 210 prompts, seed 0, KV-cached, chat template pinned); k=512 buckets with glue {59, 403} excluded per F-11; steer=bucket, mech=posterior (D-10), gate=disagree; BOTH D-11 free variants (soft = concept-bucket masses of last verified p, sharp = one-hot at its argmax) x beta {1,2,3,4}, plus the pre-registered always-gate control (sharp beta=3); paired per-prompt deltas vs the vanilla re-run at identical seeds (D-6), plug-in pooled estimates with prompt-bootstrap CI95 (seed 0, n_boot 20000); confirmatory arm named in advance: sharp beta=3. All 12 arms (11 sweep + the cover90 re-test) ran code commit 6a468f1 (vanilla clean-tree; later arms' -dirty flags are results artifacts, and the cover90 arm's additionally uncommitted docs/tools edits — src/ and tests/ clean throughout, verified by git status before commit)

Source
  • results/real/phase0b_analysis.json
  • results/real/phase0b_vanilla_k512.summary.json
  • results/real/phase0b_sharp_b3.summary.json
  • results/real/phase0b_sharp_b3_always.summary.json
  • results/real/phase0b_sharp_b3_nokv.summary.json
  • results/real/phase0b_sharp_b3_cover90.summary.json
  • results/real/glue_k512_cover90.json
  • results/real/bucket_persistence_k512.json
  • tools/analyze_phase0b.py
  • tools/bucket_persistence.py
  • docs/BLUEPRINT.md
  • docs/WHY_THIS_DIES.md
sideband/FINDINGS.md:156 · d147e8c0cd6b
Date
2026-08-30 sideband/FINDINGS.md:154 · d147e8c0cd6b
Index headline
best +0.0025 [-0.0028, +0.0079] (soft beta=1); confirmatory sharp beta=3 -0.0098 [-0.0155, -0.0039]; vanilla 0.7628; gate open 93.6-94.0%; persistence 0.027
0.7628is ZERO — no arm beats vanilla (acceptance 0.7628 pooled): best deltasideband/FINDINGS.md:157 · d147e8c0cd6b
+0.0025 [-0.0028, +0.0079](acceptance 0.7628 pooled): best delta +0.0025 [-0.0028, +0.0079] (soft beta=1), CIsideband/FINDINGS.md:157 · d147e8c0cd6b
-0.0098 [-0.0155, -0.0039](sharp beta=3) is definitively NEGATIVE, -0.0098 [-0.0155, -0.0039]; harm is dose-monotonesideband/FINDINGS.md:157 · d147e8c0cd6b
+0.0009harm is dose-monotone in beta for sharp (+0.0009 / -0.0040 / -0.0098 /sideband/FINDINGS.md:157 · d147e8c0cd6b
-0.0040dose-monotone in beta for sharp (+0.0009 / -0.0040 / -0.0098 / -0.0237sideband/FINDINGS.md:157 · d147e8c0cd6b
-0.0098in beta for sharp (+0.0009 / -0.0040 / -0.0098 / -0.0237 [-0.0302,sideband/FINDINGS.md:157 · d147e8c0cd6b
-0.0237 [-0.0302, -0.0172]for sharp (+0.0009 / -0.0040 / -0.0098 / -0.0237 [-0.0302, -0.0172]) andsideband/FINDINGS.md:157 · d147e8c0cd6b
+0.0025and null-to-slightly-negative for soft (+0.0025 / +0.0002 / -0.0051 /sideband/FINDINGS.md:157 · d147e8c0cd6b
sideband/FINDINGS.md:153 · d147e8c0cd6b sideband/docs/FINDINGS_INDEX.md:113 · aa88354d13cb
F-13Coarse gating fixes the gate, not the mechanismsidebandindex: standsNEGATIVECONFIRMEDSTANDS

Coarsening restores the gate's selectivity but every arm still loses to vanilla: the free-information actuator is dead as built at every tested granularity

Heading, as written Coarse gating fixes the gate, not the mechanism: OQ-12 answered NEGATIVE; free-information bucket steering is dead at every tested granularity

ALL SIX ARMS LAND IN THE REGISTERED BROKEN CELL — no WIN, no RESCUED-TO-NEUTRAL; OQ-12 is answered NEGATIVE by the pre-registered existential criterion, and stage 2 never triggers. sideband/FINDINGS.md:164 · d147e8c0cd6b

In plain words
NEGATIVE
the interval sits entirely below zero
CONFIRMED
a recompute from committed evidence reproduced the value
STANDS
the verdict is unchanged by later entries
index: stands
the registered verdict is unchanged and no later entry narrows it
Regime
REAL MODELS (Phase 0c stage 1) — same pair/regime as F-12 (NF4 8B + bf16 1B, gamma=4, temp 1.0, 210 committed prompts, seed 0, KV-cached); COARSE gating spaces k in {16, 32, 64} (fresh bucketize_vocab, same PCA/seed; glue by the K-7 rule verbatim: 1/1/2 buckets), all three pre-declared VIABLE (concept-cross CI lower 0.303/0.410/0.522 >= 0.25 on committed records, results/real/oq12_diagnostics.json); … sideband/FINDINGS.md:162 · d147e8c0cd6b
The full regime line

REAL MODELS (Phase 0c stage 1) — same pair/regime as F-12 (NF4 8B + bf16 1B, gamma=4, temp 1.0, 210 committed prompts, seed 0, KV-cached); COARSE gating spaces k in {16, 32, 64} (fresh bucketize_vocab, same PCA/seed; glue by the K-7 rule verbatim: 1/1/2 buckets), all three pre-declared VIABLE (concept-cross CI lower 0.303/0.410/0.522 >= 0.25 on committed records, results/real/oq12_diagnostics.json); arms = {soft, sharp} x beta=3 per k, paired against phase0b_vanilla_k512 under the gate-space exemption (buckets/glue sha differ BY DESIGN — the treatment; soundness verified on both axes: cross-k audit identity + F-11 bit-identity); protocol, verdict cells, and criterion estimand pre-registered in commit 89173bd BEFORE any steered coarse run; all six arms ran code commit 8f2f1ba (later arms -dirty by untracked sibling results only; src/tests clean throughout)

Source
  • results/real/phase0c_analysis.json
  • results/real/phase0c_k16_soft_b3.summary.json
  • results/real/phase0c_k16_sharp_b3.summary.json
  • results/real/phase0c_k32_soft_b3.summary.json
  • results/real/phase0c_k32_sharp_b3.summary.json
  • results/real/phase0c_k64_soft_b3.summary.json
  • results/real/phase0c_k64_sharp_b3.summary.json
  • results/real/oq12_diagnostics.json
  • docs/BLUEPRINT.md
  • tools/analyze_phase0b.py
sideband/FINDINGS.md:163 · d147e8c0cd6b
Date
2026-08-30 sideband/FINDINGS.md:161 · d147e8c0cd6b
Index headline
gate-open 0.55-0.56 (C=15), 0.67 (C=31), 0.80-0.81 (C=62) vs 0.94 (C=510); k16 soft -0.0029 [-0.0088, +0.0029] (the only straddle); k64 sharp -0.0252 [-0.0317, -0.0186]
0.55-0.56rate is monotone in concept-space size — 0.55-0.56 at C=15, 0.67 at C=31,sideband/FINDINGS.md:164 · d147e8c0cd6b
0.67in concept-space size — 0.55-0.56 at C=15, 0.67 at C=31, 0.80-0.81 atsideband/FINDINGS.md:164 · d147e8c0cd6b
0.80-0.81size — 0.55-0.56 at C=15, 0.67 at C=31, 0.80-0.81 at C=62, joining 0.94sideband/FINDINGS.md:164 · d147e8c0cd6b
0.940.67 at C=31, 0.80-0.81 at C=62, joining 0.94 at C=510 (F-12) — sosideband/FINDINGS.md:164 · d147e8c0cd6b
-0.0029 [-0.0088, +0.0029]acceptance, plugin [CI95]) — k16 soft -0.0029 [-0.0088, +0.0029] (the only straddle; itssideband/FINDINGS.md:164 · d147e8c0cd6b
0.559+0.0029] (the only straddle; its gate 0.559 misses the registeredsideband/FINDINGS.md:164 · d147e8c0cd6b
0.5its gate 0.559 misses the registered 0.5 bar), k16 sharp -0.0124sideband/FINDINGS.md:164 · d147e8c0cd6b
-0.0124 [-0.0186, -0.0061]misses the registered 0.5 bar), k16 sharp -0.0124 [-0.0186, -0.0061], k32 soft -0.0126sideband/FINDINGS.md:164 · d147e8c0cd6b
sideband/FINDINGS.md:160 · d147e8c0cd6b sideband/docs/FINDINGS_INDEX.md:114 · aa88354d13cb
F-14The oracle ceiling at coarse k is ~zerosidebandindex: standsSTRADDLEZEROOPEN

A perfect per-position forecast converts to at most +0.007 acceptance through the bucket-bias actuator at coarse k; registered verdict STRADDLE

Heading, as written The oracle ceiling at coarse k is ~zero: perfect forecasting barely moves acceptance (<= +0.007) through the bucket-bias actuator (attribution rider: STRADDLE, with the ceiling bounded)

REGISTERED VERDICT: STRADDLE — delta acceptance +0.0014 [-0.0043, +0.0069] (plugin, paired pooled; analytic concordance +0.0033 [-0.0010, +0.0075], positive-leaning at every offset) — attribution between forecast quality and the actuator stays formally open, reported as such per the rule. sideband/FINDINGS.md:171 · d147e8c0cd6b

In plain words
STRADDLE
the interval straddles zero
ZERO
no measurable effect; the measured value is zero within its interval
OPEN
left open; no verdict was reached
index: stands
the registered verdict is unchanged and no later entry narrows it
Regime
REAL MODELS (Phase 0c attribution rider, registered with its decision rule in commit 814fc34 BEFORE the arm existed) — oracle_nextp at k=16: per drafted position the steering posterior is the concept posterior of the TARGET'S OWN fresh p_j (target stepped incrementally against its KV cache; D-2-NAMED NON-CAUSAL CEILING, bounds only); soft mech, beta=3, disagree gate, k=16 space + committed glue; … sideband/FINDINGS.md:169 · d147e8c0cd6b
The full regime line

REAL MODELS (Phase 0c attribution rider, registered with its decision rule in commit 814fc34 BEFORE the arm existed) — oracle_nextp at k=16: per drafted position the steering posterior is the concept posterior of the TARGET'S OWN fresh p_j (target stepped incrementally against its KV cache; D-2-NAMED NON-CAUSAL CEILING, bounds only); soft mech, beta=3, disagree gate, k=16 space + committed glue; same pair/prompts/seed/NF4/KV-cached as all Phase 0b/0c arms; paired vs phase0b_vanilla_k512 under the gate-space exemption; implementation gate passed before running (beta=0 oracle path reproduces the standard path bit-exactly on the tiny harness, 3 seeds x both cache modes; steered KV A/B holds; per-position-freshness pinned by spy test)

Source
  • results/real/phase0c_oracle_analysis.json
  • results/real/phase0c_k16_oracle_b3.summary.json
  • docs/BLUEPRINT.md
  • src/sideband/real/hf_specdec.py
  • tests/test_tools_and_real.py
sideband/FINDINGS.md:170 · d147e8c0cd6b
Date
2026-08-30 sideband/FINDINGS.md:168 · d147e8c0cd6b
Index headline
+0.0014 [-0.0043, +0.0069]; truth-informed gate opens at 15.9% of steerable positions (n=69,192)
+0.007forecasting barely moves acceptance (<= +0.007) through thesideband/FINDINGS.md:167 · d147e8c0cd6b
+0.0014 [-0.0043, +0.0069]VERDICT: STRADDLE — delta acceptance +0.0014 [-0.0043, +0.0069] (plugin, paired pooled;sideband/FINDINGS.md:171 · d147e8c0cd6b
+0.0033 [-0.0010, +0.0075]paired pooled; analytic concordance +0.0033 [-0.0010, +0.0075], positive-leaning atsideband/FINDINGS.md:171 · d147e8c0cd6b
+0.0069demonstrably works, converts to at most +0.0069 acceptance (CI upper;sideband/FINDINGS.md:171 · d147e8c0cd6b
+0.0014at most +0.0069 acceptance (CI upper; point +0.0014) — the actuator'ssideband/FINDINGS.md:171 · d147e8c0cd6b
15.9%— the truth-informed gate opens at only 15.9% of steerable positionssideband/FINDINGS.md:171 · d147e8c0cd6b
84%with the target's fresh concept profile ~84% of the time, so almostsideband/FINDINGS.md:171 · d147e8c0cd6b
16%— at beta=3 soft, exploiting those 16% converts to at mostsideband/FINDINGS.md:171 · d147e8c0cd6b
sideband/FINDINGS.md:167 · d147e8c0cd6b sideband/docs/FINDINGS_INDEX.md:115 · aa88354d13cb
F-15K-2 survived at source S1: compute-tier weight-slot telemetry carries CI-positive future-concept information …sidebandindex: standsYES

Compute-tier weight-slot telemetry predicts the concept of tokens 2-3 ahead better than a probe on all 657 logged free summaries of p_a; K-2's existence question is answered YES for S1

Heading, as written K-2 survived at source S1: compute-tier weight-slot telemetry carries CI-positive future-concept information beyond the strongest free baseline at offsets 2–3, replicated on fresh harvest seeds

TELEMETRY EARNS EXISTENCE AT S1 AND THE REGISTERED REPLICATION CONFIRMS IT — the provisional qualifier lifts. sideband/FINDINGS.md:178 · d147e8c0cd6b

In plain words
index: stands
the registered verdict is unchanged and no later entry narrows it
Regime
REAL MODELS (Phase 3, source S1) — probes on the committed S1 shadow harvest (Phase 2 protocol: NF4 Llama-3.1-8B-Instruct, shadow taps at layers {7,15,23,31} = top-32-of-512 weight-slot occupancy per layer from the anchor forward only; block-start alignment, offset 1 = the token sampled from p_a itself, K-2 reads offsets >= 2; pinned prompt split 150/30/30). sideband/FINDINGS.md:176 · d147e8c0cd6b
The full regime line

REAL MODELS (Phase 3, source S1) — probes on the committed S1 shadow harvest (Phase 2 protocol: NF4 Llama-3.1-8B-Instruct, shadow taps at layers {7,15,23,31} = top-32-of-512 weight-slot occupancy per layer from the anchor forward only; block-start alignment, offset 1 = the token sampled from p_a itself, K-2 reads offsets >= 2; pinned prompt split 150/30/30). ORIGINAL grid: seeds 0..7, 1,693,036 pairs, 156 registered cells + 6 R2 shuffle, commit fba1fb3-dirty. REPLICATION (registered in BLUEPRINT with its decision PINNED to the L1/off2/linear triple BEFORE any replication artifact existed): fresh generation seeds 8..11 into the separate dir harvest_shadow_r (849,536 pairs — half volume BY REGISTERED DESIGN; the harvest tool's 1M exit-bar print and the summary's EXIT_BAR_NOT_MET verdict are the Phase-2 FULL-harvest rule, superseded by the replication registration's 4-seed pin), same 156-cell grid, commit 86136c8-dirty (both runs' -dirty = untracked results/scratch only, operator-attested; tracked src/tests clean). Riders, both runs: R1 harvest fidelity PASS bound to exact npz shas (replication: 16/16 fresh triples at rng-seed 2 first attempt, calibrated tolerances, cycle bound not stressed); R2 label-shuffle PASS at every offset (relative band; replication gaps <= 1e-4); R3 clean. Label spaces L1 concept_k16 / L2 concept_k512 / L3 concept_groups_v1 (fixed committed artifact, fit on original train split only); features (a) = ALL logged free summaries of p_a (657 dims), (b) = slots multi-hot (2048), (c) = both, (h) = h_a (4096, OQ-6 denominator); capacities linear + mlp256 per the registered economy; train seeds {0,1,2}; estimand = per-prompt paired successes summed over train seeds, then ONE prompt-cluster bootstrap CI95 over the 30 test prompts (sideband.stats.paired_cluster_delta_ci, n_boot 2000, seed 0). Archival duty discharged: second on-disk copies of both harvests sha-verified (8/8 and 4/4; operator-attested — the verification ran against the committed metas but no artifact records it).

Source
  • results/real/probes_s1.json
  • results/real/probes_s1.json.state.jsonl
  • results/real/probes_s1_replication.json
  • results/real/probes_s1_replication.json.state.jsonl
  • results/real/harvest_shadow_r/ (metas; npz sha-bound)
  • results/real/harvest_shadow_r_summary.json
  • results/real/harvest_fidelity_rider_r.json
  • results/real/harvest_shadow_summary.json
  • results/real/harvest_fidelity_rider.json
  • results/real/concept_groups_v1.json
  • results/real/harvest_audit.json
  • docs/BLUEPRINT.md
  • tools/run_probes.py
  • src/sideband/probes.py
  • src/sideband/stats.py
sideband/FINDINGS.md:177 · d147e8c0cd6b
Date
2026-08-31 sideband/FINDINGS.md:175 · d147e8c0cd6b
Index headline
pinned triple L1_concept_k16/off2/linear replication +0.0141 [+0.0081, +0.0203]; margins +0.008..+0.044 top-1 over floors 0.41-0.59; 0/4 offset-4 triples re-fired
10/12provisional qualifier lifts. Original grid: 10/12 offset->=2 (space,sideband/FINDINGS.md:178 · d147e8c0cd6b
+0.0050, all three train seeds positive, point >= +0.005); all 12 CI-positive.sideband/FINDINGS.md:178 · d147e8c0cd6b
+0.0141 [+0.0081, +0.0203]L1_concept_k16/off2/linear RE-FIRED — +0.0141 [+0.0081, +0.0203], per-seedsideband/FINDINGS.md:178 · d147e8c0cd6b
+0.0123— +0.0141 [+0.0081, +0.0203], per-seed +0.0123/+0.0144/+0.0157 — andsideband/FINDINGS.md:178 · d147e8c0cd6b
8/8EIGHT offset-2/offset-3 triples re-fired (8/8 across all three spacessideband/FINDINGS.md:178 · d147e8c0cd6b
0.006with margins agreeing across runs within 0.006 (median ~0.002: L1 off2sideband/FINDINGS.md:178 · d147e8c0cd6b
0.002agreeing across runs within 0.006 (median ~0.002: L1 off2 linear +0.0163sideband/FINDINGS.md:178 · d147e8c0cd6b
+0.0163within 0.006 (median ~0.002: L1 off2 linear +0.0163 -> +0.0141; L1 off2sideband/FINDINGS.md:178 · d147e8c0cd6b
sideband/FINDINGS.md:174 · d147e8c0cd6b sideband/docs/FINDINGS_INDEX.md:116 · aa88354d13cb
F-16The tap margin is not topic-statesidebandindex: standsP1-SURVIVESREJECTEDPASSESSURVIVES

The S1 margin is not prefix-topic state: histograms of already-generated tokens predict about nothing and the margin passes the equal-dims noise control

Heading, as written The tap margin is not topic-state: prefix-history features predict ~nothing, add ~nothing to free, and the telemetry margin passes the equal-dims noise control intact (post-hoc P1, both harvests)

MECHANICAL LABEL: P1-SURVIVES, on both harvests. sideband/FINDINGS.md:188 · d147e8c0cd6b

In plain words
P1-SURVIVES
the prefetch readout cleared its bar
PASSES
the pre-registered bar was passed
SURVIVES
the margin cleared its floor under the registered rule
index: stands
the registered verdict is unchanged and no later entry narrows it
Regime
REAL MODELS (post-hoc diagnostic P1; NON-CRITERION by registration — cannot amend F-15's verdict) — protocol registered in BLUEPRINT's P1 block BEFORE any diagnostic probe ran, with the budget-dilution amendment (equal-dims (atn) noise arm + amended DEFLATES gate) registered mid-flight on critic review BEFORE any result was read. sideband/FINDINGS.md:186 · d147e8c0cd6b
The full regime line

REAL MODELS (post-hoc diagnostic P1; NON-CRITERION by registration — cannot amend F-15's verdict) — protocol registered in BLUEPRINT's P1 block BEFORE any diagnostic probe ran, with the budget-dilution amendment (equal-dims (atn) noise arm + amended DEFLATES gate) registered mid-flight on critic review BEFORE any result was read. Feature set (t) = 560 decoder-free dims from tokens[row, :anchor] ONLY (strictly before the offset-1 token — the anchor information set; boundary pinned by live-verified mutant tests): normalized k512-alphabet concept histogram incl. merged class (511) + k16 full-prefix (16) + last-16/last-64 k16 windows (32) + log1p(len) (1). Grid: {L1, L2, L3} x offsets {2, 3} (the F-15-established horizon) x {t, at, abt, atn} x LINEAR x train seeds {0,1,2} = 72 cells per harvest, both harvests (primary seeds 0..7; replication 8..11 as no-gate confirm); (atn) = (a)+(t)+alignment-destroyed slots (slot vector of anchor (a+17) mod row-length — per-row slot statistics preserved, alignment destroyed, dims equal to (abt) at 3265; the registered amendment text wrote 3266/1218 with the pre-correction t=561 — correct values 3265/1217, dated erratum in the BLUEPRINT P1 OUTCOME). Same estimand/bootstrap as Phase 3; M3/M4/(t)-(b) pair new cells against the COMMITTED Phase-3 state-file cluster counts (load-time gate recomputes the committed payload margins from those cells at 1e-12 before anything trains). Runner commit 1a54fb7-dirty both runs (untracked results/scratch only, operator-attested).

Source
  • results/real/posthoc_topic_s1.json
  • results/real/posthoc_topic_s1.json.state.jsonl
  • results/real/posthoc_topic_s1_r.json
  • results/real/posthoc_topic_s1_r.json.state.jsonl
  • tools/run_posthoc_topic.py
  • src/sideband/probes.py
  • docs/BLUEPRINT.md
  • results/real/probes_s1.json.state.jsonl
  • results/real/probes_s1_replication.json.state.jsonl
sideband/FINDINGS.md:187 · d147e8c0cd6b
Date
2026-08-31 sideband/FINDINGS.md:185 · d147e8c0cd6b
Index headline
(t) alone max abs(acc - floor) 4.3e-4; M2 L1/off2 +0.0177 [+0.0119, +0.0232]; measured dilution M2noise -0.0045..+3.3e-5
4.3e-4floor on both harvests (max |acc - floor| = 4.3e-4 primary, 7.6e-4sideband/FINDINGS.md:188 · d147e8c0cd6b
7.6e-4(max |acc - floor| = 4.3e-4 primary, 7.6e-4 replication) — thesideband/FINDINGS.md:188 · d147e8c0cd6b
2-3accepted predicts ~nothing about concepts 2-3 tokens ahead, exactlysideband/FINDINGS.md:188 · d147e8c0cd6b
0.027what F-12's measured concept persistence (0.027, run length ~1)sideband/FINDINGS.md:188 · d147e8c0cd6b
-0.0008..+0.0042ADDS ~NOTHING TO FREE: M3 = (at)-(a) spans -0.0008..+0.0042 primary (only L3 off2sideband/FINDINGS.md:188 · d147e8c0cd6b
+0.0042 [+0.0008, +0.0077]primary (only L3 off2 CI-positive at +0.0042 [+0.0008, +0.0077], below the +0.005sideband/FINDINGS.md:188 · d147e8c0cd6b
+0.005at +0.0042 [+0.0008, +0.0077], below the +0.005 floor), max |M3| 0.0018sideband/FINDINGS.md:188 · d147e8c0cd6b
0.0018+0.0077], below the +0.005 floor), max |M3| 0.0018 replication. (iii) THEsideband/FINDINGS.md:188 · d147e8c0cd6b
sideband/FINDINGS.md:184 · d147e8c0cd6b sideband/docs/FINDINGS_INDEX.md:117 · aa88354d13cb
F-17K-2 survived at source S3: pure address-stream MoE-router telemetry carries CI-positive future-concept …sidebandindex: standsDEADYES

Pure address-stream router ids predict the concept 2 tokens ahead (3-4 under coarse labels) beyond the free baseline; K-2 answered YES for S3

Heading, as written K-2 survived at source S3: pure address-stream MoE-router telemetry carries CI-positive future-concept information beyond the strongest free baseline (offset 2 in every space; to offsets 3-4 under coarse-graining), replicated on fresh harvest seeds

TELEMETRY EARNS EXISTENCE AT S3 AND THE PINNED REPLICATION CONFIRMS IT — the provisional qualifier lifts. sideband/FINDINGS.md:197 · d147e8c0cd6b

In plain words
DEAD
the line of work is closed
index: stands
the registered verdict is unchanged and no later entry narrows it
Regime
REAL MODELS (Phase 3, source S3) — probes on the committed S3 MoE-router harvest: allenai/OLMoE-1B-7B-0125-Instruct, NF4 double-quant with routers+lm_head excluded from quantization (llm_int8_skip_modules leaf-name matched; verify_load_regime post-load inspection writes the regime into every meta from inspection, not flags); … sideband/FINDINGS.md:195 · d147e8c0cd6b
The full regime line

REAL MODELS (Phase 3, source S3) — probes on the committed S3 MoE-router harvest: allenai/OLMoE-1B-7B-0125-Instruct, NF4 double-quant with routers+lm_head excluded from quantization (llm_int8_skip_modules leaf-name matched; verify_load_regime post-load inspection writes the regime into every meta from inspection, not flags); telemetry = the model's own top-8-of-64 expert ids at each of the 16 MoE gates, anchor forward only (sorted int16, npz key "slots", 64-expert alphabet) — ADDRESSES ONLY, no router logits, no probabilities, no values; block-start alignment, offset 1 = the token sampled from p_a itself, K-2 reads offsets >= 2; pinned prompt split 150/30/30 clusters (210 prompts). PRIMARY grid: generation seeds 0..7, 1,647,264 pairs (volume bar met; splits 1,174,184/235,460/237,620), 156 registered cells + 6 R2 shuffle, partial=false, commit 808266e-dirty. REPLICATION (pinned BEFORE the primary grid existed — stronger timing than S1's, which pinned after seeing which triples fired; gate = L1_olmoe_k16/off2/linear re-fires all three conditions, non-firing branches verbatim from the S1 replication registration): fresh seeds 8..11 into the separate dir harvest_moe_r, 827,324 pairs (the tool's EXIT_BAR_NOT_MET verdict is the Phase-2 FULL-harvest rule, superseded by the registration's 4-seed pin; ran with --skip-preflight over a Windows pending-reboot advisory, flag recorded in meta args), commit 83406b1-dirty (all -dirty = untracked-only, operator-attested). Riders, both harvests: R1-S3 fidelity 16/16 PASS first attempt at rng-seed 0 / 1 (registered lineage), calibrated tolerances (expert overlap >= 5/8, mass L-inf 0.20, entropy 0.50, argmax-gap 0.20), worst observed overlap 7/8 on both, worst mass L-inf 5.57e-02 / 2.62e-02 vs calibration band 0.0666; R2 label-shuffle PASS at every offset on both runs (primary gaps <= 1.7e-5, replication <= 3.7e-4); audits clean (contamination test-in-train distinct/occurrence 0.52%/0.64% primary, 0.37%/0.47% replication; loopy rows 0 on both). Label spaces L1_olmoe_k16 / L2_olmoe_k512 (concept + merged glue class 510) / L3_olmoe_groups_v1 (fixed committed artifact, fit on the primary train chain only, sha-identical for both grids, never refit); features (a) = ALL logged free summaries of p_a, (b) = expert multi-hot (16 x 64 = 1024 dims), (c) = both, (h) = 2048-dim hidden reference (OQ-6 denominator); capacities linear + mlp256 per the registered economy; train seeds {0,1,2}; estimand identical to S1 (per-prompt paired successes summed over train seeds, ONE prompt-cluster bootstrap CI95 over the 30 test prompts, n_boot 2000, seed 0). Critic recomputed BOTH grids' full decisions from the state files alone (margins, CIs, per-seed points, firing sets, R2, floors, OQ-6) — bit-exact on both, zero blockers — before either verdict landed. Archival duty discharged: second on-disk copies of both harvests sha-verified 8/8 and 4/4 against the committed metas (operator-attested; no artifact records the check).

Source
  • results/real/probes_s3.json
  • results/real/probes_s3.json.state.jsonl
  • results/real/probes_s3_replication.json
  • results/real/probes_s3_replication.json.state.jsonl
  • results/real/harvest_moe/ (metas; npz sha-bound)
  • results/real/harvest_moe_r/ (metas)
  • results/real/harvest_moe_summary.json
  • results/real/harvest_moe_r_summary.json
  • results/real/harvest_audit_olmoe.json
  • results/real/harvest_audit_olmoe_r.json
  • results/real/harvest_fidelity_rider_olmoe.json
  • results/real/harvest_fidelity_rider_olmoe_r.json
  • results/real/concept_groups_olmoe_v1.json
  • results/real/buckets_olmoe_k16.json
  • results/real/buckets_olmoe_k32.json
  • results/real/buckets_olmoe_k64.json
  • results/real/buckets_olmoe_k512.json
  • results/real/glue_olmoe_k16.json
  • results/real/glue_olmoe_k512.json
  • results/real/fidelity_calibration_olmoe.json
  • docs/BLUEPRINT.md
  • tools/run_probes.py
  • src/sideband/real/moe_telemetry.py
  • src/sideband/real/fidelity_rider_moe.py
sideband/FINDINGS.md:196 · d147e8c0cd6b
Date
2026-09-01 sideband/FINDINGS.md:194 · d147e8c0cd6b
Index headline
pinned gate replication +0.0173 [+0.0133, +0.0221]; L3/off2 +0.0721 [+0.0652, +0.0792]; margins +0.006..+0.073 top-1 over floors 0.42-0.59
3-4(offset 2 in every space; to offsets 3-4 under coarse-graining),sideband/FINDINGS.md:193 · d147e8c0cd6b
7/12could not speak to). Primary grid: 7/12 offset->=2 (space,sideband/FINDINGS.md:197 · d147e8c0cd6b
+0.0050, all three train seeds positive, point >= +0.005), all 12 CI-positive:sideband/FINDINGS.md:197 · d147e8c0cd6b
+0.0147 [+0.0113, +0.0182]12 CI-positive: L1_olmoe_k16/off2 linear +0.0147 [+0.0113, +0.0182] and mlp256 +0.0200sideband/FINDINGS.md:197 · d147e8c0cd6b
+0.0200 [+0.0167, +0.0232]+0.0147 [+0.0113, +0.0182] and mlp256 +0.0200 [+0.0167, +0.0232]; L2_olmoe_k512/off2sideband/FINDINGS.md:197 · d147e8c0cd6b
+0.0206 [+0.0171, +0.0244][+0.0167, +0.0232]; L2_olmoe_k512/off2 +0.0206 [+0.0171, +0.0244] and off3 +0.0051sideband/FINDINGS.md:197 · d147e8c0cd6b
+0.0051 [+0.0036, +0.0066]+0.0206 [+0.0171, +0.0244] and off3 +0.0051 [+0.0036, +0.0066]; L3_olmoe_groups_v1/off2sideband/FINDINGS.md:197 · d147e8c0cd6b
+0.0721 [+0.0652, +0.0792][+0.0036, +0.0066]; L3_olmoe_groups_v1/off2 +0.0721 [+0.0652, +0.0792], off3 +0.0419 [+0.0338,sideband/FINDINGS.md:197 · d147e8c0cd6b
sideband/FINDINGS.md:193 · d147e8c0cd6b sideband/docs/FINDINGS_INDEX.md:118 · aa88354d13cb
F-18The S3 address-stream margin is not prefix-topic state at the histogram levelsidebandindex: standsP1-SURVIVESREJECTED

The S3 margin is not prefix-topic state at the histogram level; the one real topic component (L3 groups) is about a tenth of the tap's margin there

Heading, as written The S3 address-stream margin is not prefix-topic state at the histogram level: P1-SURVIVES on both harvests — topic features close ~none of the free gap at token/concept level, the router margin passes the equal-dims control intact, and the one real topic component (L3 groups) is ~1/10 the tap's margin there (post-hoc P1-analog at S3)

MECHANICAL LABEL: P1-SURVIVES on the primary harvest by the pinned rule; the replication harvest's identical mechanical label is confirmatory color only (the pinned rule gates on the primary harvest alone). sideband/FINDINGS.md:207 · d147e8c0cd6b

In plain words
P1-SURVIVES
the prefetch readout cleared its bar
index: stands
the registered verdict is unchanged and no later entry narrows it
Regime
REAL MODELS (post-hoc diagnostic P1-analog at S3; NON-CRITERION by registration — cannot amend F-17's verdict) — the conditional pin in the S3 registration (2026-08-31) fired at F-17; … sideband/FINDINGS.md:205 · d147e8c0cd6b
The full regime line

REAL MODELS (post-hoc diagnostic P1-analog at S3; NON-CRITERION by registration — cannot amend F-17's verdict) — the conditional pin in the S3 registration (2026-08-31) fired at F-17; instantiation registered 2026-09-01 in the BLUEPRINT "S3 P1-ANALOG INSTANTIATION" block AFTER F-17 landed and BEFORE any registered S3 P1 cell ran, with a dated in-block disclosure of the guard-permitted smoke (12 L1/off2 arms to a scratch out path, numbers visible before the block committed; every decision element derives from the committed conditional pin plus the committed grid payloads' provenance, none from the smoke; the smoke was later verified bit-exact against the registered primary state — an accidental determinism check). Instrument: tools/run_posthoc_topic.py source-parametrized (every new argument defaults to the committed S1 instrument — defaults asserted by tests, full-path neutrality critic-verified at the pre-run gate; a default rerun reproduces the committed S1 NUMBERS but not bytes, provenance gains feature_dims_recorded + args keys; bind_spaces_to_harvest + --n-slots binding on the load path; a cross-source primary lookup returns INCOMPLETE, never a silent verdict). Feature set (t) = 560 decoder-free dims from tokens[row, :anchor] ONLY, over the OLMoE alphabets (dims recorded in payload feature_dims_recorded, none asserted: the OLMoE alphabets merge to the same 511 k512-concepts-incl.-glue / 16 as S1, so t = 560 again; only the slots block differs — 1024 = 16 gates x 64 experts). Grid: {L1,L2,L3}_olmoe x offsets {2,3} x {t, at, abt, atn} x LINEAR x train seeds {0,1,2} = 72 cells per harvest, BOTH committed S3 harvests (primary seeds 0..7, runtime 14,265.9 s; replication seeds 8..11, 7,202.9 s; payload runtime >= sum of cell runtimes on both — fresh single passes, no resumed cells); (atn) = (a)+(t)+alignment-destroyed expert vectors (anchor (a+17) mod row-length), dims equal to (abt) by construction. Same estimand/bootstrap as Phase 3 (seed-summed prompt-cluster bootstrap, n_boot 2000, seed 0, 30 test clusters); M3/M4/(t)-(b) pair against the COMMITTED S3 grid state files via their sha headers + the 1e-12 payload-reproduction gate before anything trains. Runner commit f3bb32e-dirty on both payloads, where the dirt IS the instrument + registration that ran (reviewed at the pre-run critic gate, committed together with this finding; the smoke bit-exactness and the recompute make silent drift immaterial to the numbers). Critic recomputed ALL SIX readout families and BOTH conclusion labels from the state files alone (fresh implementation, zero imports from the repo): max abs deviation 0.000e+00 across every margin, CI bound, and per-seed point; grid identity exact (1 sha header + 72 unique cells each, 30 clusters per cell, state order == grid order); sha chain closed (posthoc state headers == committed grid headers == harvest metas; phase3 and label-space shas match disk and the grid payloads); ZERO BLOCKERS.

Source
  • results/real/posthoc_topic_s3.json
  • results/real/posthoc_topic_s3.json.state.jsonl
  • results/real/posthoc_topic_s3_r.json
  • results/real/posthoc_topic_s3_r.json.state.jsonl
  • results/real/probes_s3.json.state.jsonl
  • results/real/probes_s3_replication.json.state.jsonl
  • tools/run_posthoc_topic.py
  • tests/test_probes.py
  • src/sideband/probes.py
  • docs/BLUEPRINT.md
sideband/FINDINGS.md:206 · d147e8c0cd6b
Date
2026-09-01 sideband/FINDINGS.md:204 · d147e8c0cd6b
Index headline
M2 L1_olmoe_k16/off2/linear +0.012958 [+0.008783, +0.016809]; L1/L2 M3 abs(point) <= 0.00094; L3/off3 M3 +0.006972 [+0.003302, +0.011127]
1/10one real topic component (L3 groups) is ~1/10 the tap's margin theresideband/FINDINGS.md:203 · d147e8c0cd6b
+0.012958 [+0.008783, +0.016809]= (abt)-(at) at L1_olmoe_k16/off2/linear = +0.012958 [+0.008783, +0.016809], per-seedsideband/FINDINGS.md:207 · d147e8c0cd6b
+0.012667+0.012958 [+0.008783, +0.016809], per-seed +0.012667/+0.013356/+0.012852 —sideband/FINDINGS.md:207 · d147e8c0cd6b
+0.005> 0, all three seeds positive, point >= +0.005. Replication: +0.017465sideband/FINDINGS.md:207 · d147e8c0cd6b
+0.017465 [+0.013190, +0.022209]positive, point >= +0.005. Replication: +0.017465 [+0.013190, +0.022209]; row-wise |delta| vssideband/FINDINGS.md:207 · d147e8c0cd6b
0.0045+0.022209]; row-wise |delta| vs primary <= 0.0045 (M2) / <= 0.0051 (M2n).sideband/FINDINGS.md:207 · d147e8c0cd6b
0.0051|delta| vs primary <= 0.0045 (M2) / <= 0.0051 (M2n). (ii) (t) ADDSsideband/FINDINGS.md:207 · d147e8c0cd6b
0.00094rows are CI-crossing with |point| <= 0.00094 on the primary harvestsideband/FINDINGS.md:207 · d147e8c0cd6b
sideband/FINDINGS.md:203 · d147e8c0cd6b sideband/docs/FINDINGS_INDEX.md:119 · aa88354d13cb
F-19K-3 Attempt #1 NULL: adaptive block scheduling on frozen telemetry probes does not convert the S1 probe …sidebandindex: standsNULLNOT WIN

Adaptive block scheduling on the frozen S1 probes does not convert: telemetry ties vanilla in tokens per verify call, far below the registered bar; K-3 fires

Heading, as written K-3 Attempt #1 NULL: adaptive block scheduling on frozen telemetry probes does not convert the S1 probe margins into tokens-per-target-call at matched drafting budget — every gate leg fails on the fresh verdict population (telemetry vs vanilla +0.007 against a +0.062 bar, CI straddling zero), the pairwise budget guard breached twice by MECHANISM, and the information-free gamma-6 control confirms the gaming arithmetic the guard existed to block

VERDICT (seed-1 void-rerun, the registered single rerun): NOT WIN. sideband/FINDINGS.md:214 · d147e8c0cd6b

In plain words
NULL
no effect found against the bar
NOT WIN
the point estimate did not reach the win bar
index: stands
the registered verdict is unchanged and no later entry narrows it
Regime
REAL MODELS (Llama-3.1-8B-Instruct target + Llama-3.2-1B-Instruct draft, NF4, RTX 3080 Ti). The registered Phase-4 K-3 Attempt #1 arc executed end to end: arc registration + amendments #1/#2, D2/D3 diagnostics, stage-E clarifications E1-E12, tolerance amendment, threshold-search registration (+ dated erratum + robustness contingency), void execution, fold-back count amendment — … sideband/FINDINGS.md:213 · d147e8c0cd6b
The full regime line

REAL MODELS (Llama-3.1-8B-Instruct target + Llama-3.2-1B-Instruct draft, NF4, RTX 3080 Ti). The registered Phase-4 K-3 Attempt #1 arc executed end to end: arc registration + amendments #1/#2, D2/D3 diagnostics, stage-E clarifications E1-E12, tolerance amendment, threshold-search registration (+ dated erratum + robustness contingency), void execution, fold-back count amendment — every block dated before the result it governs. Arms: 1 vanilla gamma-4; 2 draft-confidence (q-based stop/extend, the incumbent-style comparator); 3 (a)-probe free-signal; 4 (c)-probe telemetry (the K-3 bet); 5 fixed gamma-6 descriptive control. Thresholds frozen on the 30 val prompts by the registered constrained-max frontier search (97 evals, all ratios critic-recomputed bit-exact from state files; selections arm 2 stop q0.25/extend q0.5, arm 3 q0.95/q0.05, arm 4 q0.4/q0.05, re-frozen q0.35/q0.05 under the void). Verdict population = data/prompts_k3_verdict.jsonl (210 fresh committed prompts, authored before any D2/D3 output was read). Estimand = pooled emitted tokens per verify call (E6); delta = stats.paired_ratio_delta_ci (joint cluster bootstrap, n_boot 2000, seed 0). WIN gate (all-hold): for EVERY comparator in {1,2,3}: pairwise delta CI95 lower > 0 AND point >= +0.062 AND arm-4 drafted-per-emitted <= comparator + 0.01.

Source
results/real/k3_verdict_s1_arm1.summary.json, results/real/k3_verdict_s1_arm2.summary.json, results/real/k3_verdict_s1_arm3.summary.json, results/real/k3_verdict_s1_arm4.summary.json, results/real/k3_verdict_s1_arm5.summary.json, results/real/k3_verdict_s1_arm1.stats.jsonl, results/real/k3_verdict_s1_arm2.stats.jsonl, …
The full source line

results/real/k3_verdict_s1_arm1.summary.json, results/real/k3_verdict_s1_arm2.summary.json, results/real/k3_verdict_s1_arm3.summary.json, results/real/k3_verdict_s1_arm4.summary.json, results/real/k3_verdict_s1_arm5.summary.json, results/real/k3_verdict_s1_arm1.stats.jsonl, results/real/k3_verdict_s1_arm2.stats.jsonl, results/real/k3_verdict_s1_arm3.stats.jsonl, results/real/k3_verdict_s1_arm4.stats.jsonl, results/real/k3_verdict_s1_arm5.stats.jsonl (seed-1 gate data; metas/gen alongside; byproduct sha256 bindings in summaries), results/real/k3_verdict_arm1.summary.json (seed-0 P-stage record, commit cbe3ee3), results/real/k3_verdict_arm4.summary.json, results/real/k3_verdict_arm5.summary.json (seed-0 VOIDED, commit 48814ac), results/real/k3_thresholds_arm2.json, results/real/k3_thresholds_arm3.json, results/real/k3_thresholds_arm4.json, results/real/k3_thresholds_arm4_v2.json, results/real/k3_threshold_search_arm2.json.state.jsonl, results/real/k3_threshold_search_arm3.json.state.jsonl, results/real/k3_threshold_search_arm4.json.state.jsonl (commit e0ae94f), docs/BLUEPRINT.md (arc registration, E1-E12, threshold-search registration + erratum + contingency, void execution, fold-back count amendment), docs/ECONOMICS.md (S5 pinned bars). Commit chain 69cb8f2 -> 7d3450f -> 36600cb -> 0663864 -> e0ae94f -> cbe3ee3 -> 48814ac -> 92fa9cb -> this commit.

sideband/FINDINGS.md:223 · d147e8c0cd6b
Date
2026-09-03 sideband/FINDINGS.md:212 · d147e8c0cd6b
Index headline
arm4-vs-arm1 +0.006860 [-0.015024, +0.025572] against bar +0.062; val-tuned +0.060 shrank to +0.007; vanilla 3.020427 tokens/call, telemetry 3.027287
+0.007verdict population (telemetry vs vanilla +0.007 against a +0.062 bar,sideband/FINDINGS.md:211 · d147e8c0cd6b
+0.062(telemetry vs vanilla +0.007 against a +0.062 bar, CI straddlingsideband/FINDINGS.md:211 · d147e8c0cd6b
+0.006860 CI95 [-0.015024, +0.025572]single rerun): NOT WIN. arm4-vs-arm1 delta +0.006860 CI95 [-0.015024, +0.025572]; arm4-vs-arm3 +0.006974sideband/FINDINGS.md:214 · d147e8c0cd6b
+0.006974 [-0.014909, +0.025690]CI95 [-0.015024, +0.025572]; arm4-vs-arm3 +0.006974 [-0.014909, +0.025690]; arm4-vs-arm2 +0.048863sideband/FINDINGS.md:214 · d147e8c0cd6b
+0.048863 [-0.007393, +0.104327][-0.014909, +0.025690]; arm4-vs-arm2 +0.048863 [-0.007393, +0.104327] with the budget guardsideband/FINDINGS.md:214 · d147e8c0cd6b
1.317033guard breached a SECOND time (arm 4 dpe 1.317033 vs allowed 1.285681;sideband/FINDINGS.md:214 · d147e8c0cd6b
1.285681SECOND time (arm 4 dpe 1.317033 vs allowed 1.285681; margin +0.0314, vssideband/FINDINGS.md:214 · d147e8c0cd6b
+0.03144 dpe 1.317033 vs allowed 1.285681; margin +0.0314, vs seed-0 +0.0289).sideband/FINDINGS.md:214 · d147e8c0cd6b
sideband/FINDINGS.md:211 · d147e8c0cd6b sideband/docs/FINDINGS_INDEX.md:120 · aa88354d13cb
F-20OQ-15 Stage I: the registered existence gate passes mechanically, but the registered stale-target control …sidebandindex: qualifiedARTIFACT-DOMINANTPASS

The existence gate passes mechanically, but a five-seed stale-target control leaves the margin indistinguishable from zero-information draws; the free-features head beats the telemetry head outright

Heading, as written OQ-15 Stage I: the registered existence gate passes mechanically, but the registered stale-target control (five permutation seeds) leaves the margin indistinguishable at this power from zero-information target draws (pinned label ARTIFACT-DOMINANT) — block-aligned acceptance-relevant telemetry information above free signals is NOT demonstrated, and the telemetry head loses outright to the free-features head despite its +2048 extra dims

STAGE-I GATE (mechanical): PASS. sideband/FINDINGS.md:228 · d147e8c0cd6b

In plain words
ARTIFACT-DOMINANT
a stale control with no live information reproduces at least half of the measured margin
PASS
the registered gate passed
index: qualified
the verdict stands as registered, but a later finding or dated note narrows its scope or meaning
Regime
REAL MODELS (Llama-3.1-8B-Instruct target + Llama-3.2-1B-Instruct draft, NF4, RTX 3080 Ti; S1-style value-tier slot telemetry, 4 layers x top-32). CPU-only analysis of the K-3 arm-1 (vanilla gamma-4) byproduct logs, both verdict seeds (17,673 + 17,581 blocks x 4 drafted offsets) + val/tapcal runs as TRAIN-only rows. sideband/FINDINGS.md:227 · d147e8c0cd6b
The full regime line

REAL MODELS (Llama-3.1-8B-Instruct target + Llama-3.2-1B-Instruct draft, NF4, RTX 3080 Ti; S1-style value-tier slot telemetry, 4 layers x top-32). CPU-only analysis of the K-3 arm-1 (vanilla gamma-4) byproduct logs, both verdict seeds (17,673 + 17,581 blocks x 4 drafted offsets) + val/tapcal runs as TRAIN-only rows. Registered pre-run (OQ-15 FOLD-BACK #1 ARC REGISTRATION, commit 89b38ef): prompt-level split default_rng(15) -> 126/42/42, target = per-position ANALYTIC acceptance 1 - TV(p_j, q_j) (D-9 headline form, not survival-conditioned), features via the committed scheduler.probe_input (FREE 657 dims + q_sig availability rule; TEL +2048 slot multi-hot; NOISE = F-16's alignment-destroying shift), closed-form ridge with lambda-on-VAL the only DOF, gate = delta_noise = pooled TEST MSE(NOISE) - MSE(TEL) via stats.paired_cluster_delta_ci (42 prompt clusters pooling both seed-runs, 28,616 positions, n_boot 2000, seed 0). The DI-1 stale-target control was critic-proposed at the recompute round and REGISTERED with pinned conclusion labels BEFORE it ran (the P1 post-hoc pattern).

Source
results/real/oq15_stage1.json, results/real/oq15_stage1.json.state.jsonl, results/real/oq15_stage1.json.meta.json (gate run), results/real/oq15_stage1_rotcheck.json, results/real/oq15_stage1_rotcheck.json.state.jsonl, results/real/oq15_stage1_rotcheck.json.meta.json (DI-1 primary), …
The full source line

results/real/oq15_stage1.json, results/real/oq15_stage1.json.state.jsonl, results/real/oq15_stage1.json.meta.json (gate run), results/real/oq15_stage1_rotcheck.json, results/real/oq15_stage1_rotcheck.json.state.jsonl, results/real/oq15_stage1_rotcheck.json.meta.json (DI-1 primary), results/real/oq15_stage1_rotcheck_r.json, results/real/oq15_stage1_rotcheck_r.json.state.jsonl, results/real/oq15_stage1_rotcheck_r.json.meta.json (DI-1 replicate), results/real/oq15_stage1_rotcheck_s2.json, results/real/oq15_stage1_rotcheck_s3.json, results/real/oq15_stage1_rotcheck_s4.json (seed-spread extension; states/metas alongside), tools/train_outcome_head.py, src/sideband/outcome_head.py, docs/BLUEPRINT.md (OQ-15 FOLD-BACK #1 ARC REGISTRATION + DI-1 registration + outcome notes). Commit chain 89b38ef -> this commit.

sideband/FINDINGS.md:234 · d147e8c0cd6b
Date
2026-09-03 sideband/FINDINGS.md:226 · d147e8c0cd6b
Index headline
delta_noise +0.000763 [+0.000308, +0.001215]; stale-target primary +0.000643 [-0.000076, +0.001357]; delta_free -0.00286 [-0.00322, -0.00252]
+2048to the free-features head despite its +2048 extra dimssideband/FINDINGS.md:225 · d147e8c0cd6b
+0.000763 CI95 [+0.000308, +0.001215]GATE (mechanical): PASS. delta_noise = +0.000763 CI95 [+0.000308, +0.001215]. Critic recompute fromsideband/FINDINGS.md:228 · d147e8c0cd6b
02:50Timeline clean (registration commit 02:50, run 03:35).sideband/FINDINGS.md:228 · d147e8c0cd6b
03:35clean (registration commit 02:50, run 03:35).sideband/FINDINGS.md:228 · d147e8c0cd6b
sideband/FINDINGS.md:225 · d147e8c0cd6b sideband/docs/FINDINGS_INDEX.md:121 · aa88354d13cb
F-21OQ-15 Stage C (fold-back #1 terminal)sidebandindex: standsCONVERSION-MODE NULLNOT WINNEGATIVE

The telemetry head ties vanilla and loses CI-clear to the identical policy driven by free features: CONVERSION-MODE NULL

Heading, as written OQ-15 Stage C (fold-back #1 terminal): CONVERSION-MODE NULL — the outcome-supervised telemetry head does not convert to tokens-per-call at matched budget on the fresh verdict population: it ties vanilla (+0.003, CI straddling zero) and loses CI-clear to the free-features head arm (-0.082), the live-conversion mirror of F-20's delta_free; the seed-2 guard breach was a selection-time budget gap (not inflation), resolved by the registered void semantics with the guard clearing every leg at seed 3

VERDICT (seed-3 void-rerun, the registered single rerun): NOT WIN — every leg fails BOTH statistical criteria, and NO leg fails on the guard (arm 7's dpe 1.2742347176420175 sits below all three comparators outright: 1.3296655675930287 / 1.3170482039828773 / 1.2999498131935538) — the null is not a guard technicality. sideband/FINDINGS.md:239 · d147e8c0cd6b

In plain words
CONVERSION-MODE NULL
the signal may exist, but this mechanism did not convert it into more emitted tokens per check
NOT WIN
the point estimate did not reach the win bar
NEGATIVE
the interval sits entirely below zero
index: stands
the registered verdict is unchanged and no later entry narrows it
Regime
REAL MODELS (Llama-3.1-8B-Instruct target + Llama-3.2-1B-Instruct draft, NF4, RTX 3080 Ti). The registered OQ-15 Stage-C conversion stage executed end to end under the arc registration (commit 89b38ef) + STAGE-C FREEZE amendment (9e8fa4a) + STAGE-C VOID EXECUTION (46ab439) — every block dated before the result it governs. sideband/FINDINGS.md:238 · d147e8c0cd6b
The full regime line

REAL MODELS (Llama-3.1-8B-Instruct target + Llama-3.2-1B-Instruct draft, NF4, RTX 3080 Ti). The registered OQ-15 Stage-C conversion stage executed end to end under the arc registration (commit 89b38ef) + STAGE-C FREEZE amendment (9e8fa4a) + STAGE-C VOID EXECUTION (46ab439) — every block dated before the result it governs. Arms: C1 vanilla gamma-4; C2 draft-confidence (the incumbent-style comparator, K-3 arm-2 policy retuned on fresh val); C3 = head_free (outcome-head policy on FREE features, 657 dims — the matched-actuator free baseline); C4 = head_tel (the OQ-15 bet: same policy on FREE+TEL, 2705 dims). Head policy: STOP at consult j in {1,2} iff S_j = prod a_hat_1..j < tau_stop[j] (strict); EXTEND to gamma 6 iff S_4*a_hat_5 >= tau_ext; a_hat from the frozen ridge heads — offsets 1-4 Stage-I, offset 5 trained at freeze time on the arm-5 byproducts (oq15_head_weights.npz, sha 35754b8a2b05.., binding-gated export bit-exact vs the committed Stage-I state). Fresh populations committed with the registration BEFORE Stage I ran: 30 val (searches) + 210 verdict prompts (prompts_sha 71fc9995151e..). B_ref frozen at first read = 1.3255813953488371 (oq15c_val_arm1, arm-1 val run, t/c 3.017543859649123). Thresholds by the registered constrained-max frontier search on val only (128 evals total: arm 2 SELECTED stop q0.1/extend q0.75 val t/c 3.1036761668731927 budget 1.2989087037529945, 41 evals; arm 6 q0.1/q0.75, 3.050039714058777, 1.2858072916666667, 44 evals; arm 7 q0.05/q0.9, 3.0600961538461537, 1.3119926682377585, 43 evals). Estimand = pooled emitted tokens per verify call (E6); delta = stats.paired_ratio_delta_ci (joint cluster bootstrap, n_boot 2000, seed 0). WIN gate (all-hold): for EVERY comparator in {C1, C2, C3}: pairwise delta CI95 lower > 0 AND point >= +0.062 AND C4 drafted-per-emitted <= comparator's + 0.01. Containment re-validated pre-verdict (oq15c_containment_rider.json, PASS on the strictest bit-identity branch).

Source
results/real/oq15c_verdict_s3_arm1.summary.json, results/real/oq15c_verdict_s3_arm2.summary.json, results/real/oq15c_verdict_s3_arm6.summary.json, results/real/oq15c_verdict_s3_arm7.summary.json, results/real/oq15c_verdict_s3_arm1.stats.jsonl, results/real/oq15c_verdict_s3_arm2.stats.jsonl, results/real/oq15c_verdict_s3_arm6.stats.jsonl, results/real/oq15c_verdict_s3_arm7.stats.jsonl (seed-3 gate data; …
The full source line

results/real/oq15c_verdict_s3_arm1.summary.json, results/real/oq15c_verdict_s3_arm2.summary.json, results/real/oq15c_verdict_s3_arm6.summary.json, results/real/oq15c_verdict_s3_arm7.summary.json, results/real/oq15c_verdict_s3_arm1.stats.jsonl, results/real/oq15c_verdict_s3_arm2.stats.jsonl, results/real/oq15c_verdict_s3_arm6.stats.jsonl, results/real/oq15c_verdict_s3_arm7.stats.jsonl (seed-3 gate data; metas/gen alongside; byproduct sha256 bindings in summaries), results/real/oq15c_verdict_arm1.summary.json, results/real/oq15c_verdict_arm2.summary.json, results/real/oq15c_verdict_arm6.summary.json, results/real/oq15c_verdict_arm7.summary.json (seed-2: arm-7 VOIDED, comparators superseded; commit 46ab439), results/real/oq15c_thresholds_arm2.json, results/real/oq15c_thresholds_arm6.json, results/real/oq15c_thresholds_arm7.json, results/real/oq15c_thresholds_arm7_v2.json, results/real/oq15c_search_arm2.json, results/real/oq15c_search_arm6.json, results/real/oq15c_search_arm7.json, results/real/oq15c_search_arm2.json.state.jsonl, results/real/oq15c_search_arm6.json.state.jsonl, results/real/oq15c_search_arm7.json.state.jsonl, results/real/oq15c_val_arm1.summary.json, results/real/oq15c_containment_rider.json, results/real/oq15_head_weights.npz (+ meta), results/real/oq15_stage1.json (the frozen heads' provenance), data/prompts_oq15_verdict.jsonl, data/prompts_oq15_val.jsonl, docs/BLUEPRINT.md (arc registration, STAGE-C FREEZE, STAGE-C VOID EXECUTION, STAGE-C OUTCOME). Commit chain 89b38ef -> 44b08fa -> 9e8fa4a -> 46ab439 -> this commit.

sideband/FINDINGS.md:249 · d147e8c0cd6b
Date
2026-09-04 sideband/FINDINGS.md:237 · d147e8c0cd6b
Index headline
C4-vs-C1 +0.002664 [-0.044719, +0.046727]; C4-vs-C3 -0.081846 [-0.132484, -0.035567]; vanilla 3.00827523663776 tokens/call, head_free 3.0927852831273355, head_tel 3.010939077751599
+0.003fresh verdict population: it ties vanilla (+0.003, CI straddling zero)sideband/FINDINGS.md:236 · d147e8c0cd6b
-0.082CI-clear to the free-features head arm (-0.082), the live-conversionsideband/FINDINGS.md:236 · d147e8c0cd6b
1.2742347176420175and NO leg fails on the guard (arm 7's dpe 1.2742347176420175 sits below all threesideband/FINDINGS.md:239 · d147e8c0cd6b
1.3296655675930287sits below all three comparators outright: 1.3296655675930287 / 1.3170482039828773 /sideband/FINDINGS.md:239 · d147e8c0cd6b
1.3170482039828773comparators outright: 1.3296655675930287 / 1.3170482039828773 / 1.2999498131935538) —sideband/FINDINGS.md:239 · d147e8c0cd6b
1.29994981319355381.3296655675930287 / 1.3170482039828773 / 1.2999498131935538) — the null is not asideband/FINDINGS.md:239 · d147e8c0cd6b
+0.002664 CI95 [-0.044719, +0.046727]null is not a guard technicality. C4-vs-C1 +0.002664 CI95 [-0.044719, +0.046727]; C4-vs-C2 -0.050425sideband/FINDINGS.md:239 · d147e8c0cd6b
-0.050425 [-0.109388, +0.005772]CI95 [-0.044719, +0.046727]; C4-vs-C2 -0.050425 [-0.109388, +0.005772]; C4-vs-C3 -0.081846sideband/FINDINGS.md:239 · d147e8c0cd6b
sideband/FINDINGS.md:236 · d147e8c0cd6b sideband/docs/FINDINGS_INDEX.md:122 · aa88354d13cb
F-22OQ-16 Stage R: STAGE-R PASS — the S1 telemetry stream linearly reconstructs the anchor hidden state h_a above …sidebandindex: qualifiedSTAGE-R PASSREFUTED

Telemetry linearly reconstructs the anchor state above the noise twin and the free features (Stage R PASS), but the registered magnitudes are pathology-inflated; a free hashed token encoding (FREETOK) reconstructs slightly better; telemetry plus FREETOK (UNION) beats FREETOK; …

The full index summary

Telemetry linearly reconstructs the anchor state above the noise twin and the free features (Stage R PASS), but the registered magnitudes are pathology-inflated; a free hashed token encoding (FREETOK) reconstructs slightly better; telemetry plus FREETOK (UNION) beats FREETOK; a second token comparator (FREETOK2, built to be stronger) landed CI-clear worse than FREETOK (COMPARATOR-TOOTHLESS), so "beyond p" stays unpriced

Heading, as written OQ-16 Stage R: STAGE-R PASS — the S1 telemetry stream linearly reconstructs the anchor hidden state h_a above both the alignment-destroyed twin (+0.176, ~85 stale-sd clear) and the free-features baseline (+0.205), the registered gate's first surprising branch and the refutation of its own pinned expectation; but the registered magnitudes are inflated ~25% by a standardization pathology on two out-of-support test prompts, the honest deployable margin against the mean predictor is NOT CI-clear at the pooled estimand, and the post-gate diagnostics DR-1/DR-2 (registered before running) price what the pass is worth

GATE (mechanical): STAGE-R PASS. sideband/FINDINGS.md:254 · d147e8c0cd6b

In plain words
STAGE-R PASS
the reconstruction stage passed its registered gate
REFUTED
a recompute from committed evidence contradicted the stated value
index: qualified
the verdict stands as registered, but a later finding or dated note narrows its scope or meaning
Regime
REAL MODELS (Llama-3.1-8B-Instruct target + Llama-3.2-1B-Instruct draft, NF4, RTX 3080 Ti; S1-style value-tier slot telemetry, 4 layers x top-32 hash-sketched addresses at the down_proj inputs of layers 7/15/23/31). sideband/FINDINGS.md:253 · d147e8c0cd6b
The full regime line

REAL MODELS (Llama-3.1-8B-Instruct target + Llama-3.2-1B-Instruct draft, NF4, RTX 3080 Ti; S1-style value-tier slot telemetry, 4 layers x top-32 hash-sketched addresses at the down_proj inputs of layers 7/15/23/31). CPU-only analysis of the FOUR vanilla gamma-4 verdict byproduct runs (K-3 seeds 0/1 on data/prompts_k3_verdict.jsonl; OQ-15C seeds 2/3 on data/prompts_oq15_verdict.jsonl; 70,529 blocks; runs [0] and [2] are voided attempts' vanilla arms — logging was never quarantined, disclosed at registration). Registered pre-run (OQ-16 FOLD-BACK #2 ARC REGISTRATION, commit bd71358): target = the byproduct-logged anchor hidden state h_a (model.norm output, 4096 dims, float16 on disk — the cast is part of the target), z-scored per coordinate on pooled TRAIN moments; features = anchor-time core only through the committed scheduler.probe_input (FREE 657 / TEL 2705 / NOISE = F-16 shift 17; NO q_sig — it does not exist at the anchor, critic-verified in k3_live/hf_specdec); prompt-level 126/42/42 splits per population (Stage-I recipe), cluster = (population, prompt), 84 TEST clusters; multi-output closed-form ridge, lambda the only DOF over the WIDENED grid 1e-2..1e5; estimand = pooled TEST standardized MSE, deltas by stats.paired_ratio_delta_ci (per-cluster SSE / position counts, n_boot 2000, seed 0); GATE = two-leg AND: leg 1 delta_noise CI-lower > 0 AND point > stale_mean + 3 sd AND point > stale max over a GATING five-seed stale-target control (DI-1 recipe, lambda re-selected inside every control); leg 2 delta_free CI-lower > 0. Three design decisions put to the user before the registration (all-four-vanilla-runs / two-leg gate / direction proxy reported-only). Timeline: registration commit bd71358 -> registered run (no overrides, meta bd71358-dirty from the untracked HANDOFF only) -> critic attack round -> DR-1/DR-2 registered -> DR runs -> this finding. The DR-registration-before-DR-run ordering (registration text ~3.5 min before the runs' started_unix, ~25 min before any DR result) is provable from the session transcript only; the committed record binds the DR metas in the same commit as the text — operator-attested, the -dirty pattern.

Source
results/real/oq16_stage_r.json, results/real/oq16_stage_r.json.state.jsonl, results/real/oq16_stage_r.json.meta.json (registered run), results/real/oq16_stage_r_inv.json, results/real/oq16_stage_r_inv.json.state.jsonl, results/real/oq16_stage_r_inv.json.meta.json (invariance leg), results/real/oq16_stage_r_dr.json, …
The full source line

results/real/oq16_stage_r.json, results/real/oq16_stage_r.json.state.jsonl, results/real/oq16_stage_r.json.meta.json (registered run), results/real/oq16_stage_r_inv.json, results/real/oq16_stage_r_inv.json.state.jsonl, results/real/oq16_stage_r_inv.json.meta.json (invariance leg), results/real/oq16_stage_r_dr.json, results/real/oq16_stage_r_dr.json.state.jsonl, results/real/oq16_stage_r_dr.json.meta.json (DR-1/DR-2), results/real/oq16_stage_r_dr3.json, results/real/oq16_stage_r_dr3.json.state.jsonl, results/real/oq16_stage_r_dr3.json.meta.json (DR-3), results/real/oq16_stage_r_dr4.json, results/real/oq16_stage_r_dr4.json.state.jsonl, results/real/oq16_stage_r_dr4.json.meta.json (DR-4), data/prompts_oq16_verdict.jsonl, data/make_prompts_oq16.py (Stage-A verdict population), results/real/k3_verdict_arm1.summary.json, results/real/k3_verdict_s1_arm1.summary.json, results/real/oq15c_verdict_arm1.summary.json, results/real/oq15c_verdict_s3_arm1.summary.json (byproduct sha bindings; raws gitignored), src/sideband/recon.py, tools/train_recon.py, tests/test_recon.py, docs/BLUEPRINT.md (OQ-16 ARC REGISTRATION, STAGE-R OUTCOME, DR-1/DR-2 registration + outcome, DR-3 registration + outcome, DR-4 registration + outcome, STAGE-A FREEZE), src/sideband/real/steer_inject.py, tools/train_inject.py, tools/run_k3_arms.py, tools/train_recon.py (Stage-A instrument, pre-GPU). Commit chain ad32345 -> bd71358 -> 2ec789f -> 7fb3440 -> c2181b1 -> the DR-4 outcome commit -> this commit.

sideband/FINDINGS.md:263 · d147e8c0cd6b
Date
2026-09-04 sideband/FINDINGS.md:252 · d147e8c0cd6b
Index headline
registered delta_noise +0.176469 [+0.130414, +0.251518], delta_free +0.204984 [+0.153698, +0.288162]; DR-1 tel_vs_trivial +0.250784 [+0.239316, +0.261616] (R^2 +0.253); DR-2 tel_vs_freetok -0.004792 [-0.007192, -0.002600]; DR-3 union_vs_freetok +0.065180 [+0.062411, +0.067705]; …
The full headline

registered delta_noise +0.176469 [+0.130414, +0.251518], delta_free +0.204984 [+0.153698, +0.288162]; DR-1 tel_vs_trivial +0.250784 [+0.239316, +0.261616] (R^2 +0.253); DR-2 tel_vs_freetok -0.004792 [-0.007192, -0.002600]; DR-3 union_vs_freetok +0.065180 [+0.062411, +0.067705]; DR-4 freetok2_vs_freetok -0.003085 [-0.004984, -0.001280]

+0.176above both the alignment-destroyed twin (+0.176, ~85 stale-sd clear)sideband/FINDINGS.md:251 · d147e8c0cd6b
+0.205clear) and the free-features baseline (+0.205), the registered gate'ssideband/FINDINGS.md:251 · d147e8c0cd6b
25%but the registered magnitudes are inflated ~25% by a standardizationsideband/FINDINGS.md:251 · d147e8c0cd6b
+0.176469 CI95 [+0.130414, +0.251518]STAGE-R PASS. Leg 1: delta_noise = +0.176469 CI95 [+0.130414, +0.251518]; the five stale-seedsideband/FINDINGS.md:254 · d147e8c0cd6b
+0.000343+0.251518]; the five stale-seed points +0.000343 / +0.000849 / -0.000698sideband/FINDINGS.md:254 · d147e8c0cd6b
+0.000849the five stale-seed points +0.000343 / +0.000849 / -0.000698 / -0.004252sideband/FINDINGS.md:254 · d147e8c0cd6b
-0.000698stale-seed points +0.000343 / +0.000849 / -0.000698 / -0.004252 / +0.000294sideband/FINDINGS.md:254 · d147e8c0cd6b
-0.004252points +0.000343 / +0.000849 / -0.000698 / -0.004252 / +0.000294 (meansideband/FINDINGS.md:254 · d147e8c0cd6b
sideband/FINDINGS.md:251 · d147e8c0cd6b sideband/docs/FINDINGS_INDEX.md:123 · aa88354d13cb
F-23OQ-16 CLOSED BY DATED ERRATUM (TARGET-FREE)sidebandindex: standsCLOSED BY DATED ERRATUMTARGET-FREESTANDSFREE-ACTUATOR WINNULL

The registered reconstruction target h_a is an exact linear function of the anchor's logits and hence of the full next-token distribution p, free information under D-5; the stream's reconstruction is a lossy copy of what the decoder already has; the Stage-A telemetry arms are struck pre-GPU; …

The full index summary

The registered reconstruction target h_a is an exact linear function of the anchor's logits and hence of the full next-token distribution p, free information under D-5; the stream's reconstruction is a lossy copy of what the decoder already has; the Stage-A telemetry arms are struck pre-GPU; no partition branch claimed (limbo clause honored), mechanism TARGET-FREE

Heading, as written OQ-16 CLOSED BY DATED ERRATUM (TARGET-FREE): the anchor hidden state the reconstruction lane targeted is an exact linear function of the verifier's own next-token distribution — free information under D-5 — so the telemetry reconstruction is a lossy copy of what the decoder already has and the Stage-A telemetry actuator test is void before any GPU spend; the lane is closed with no partition branch claimed (limbo clause honored)

CLAIM: h_a (the final-norm output at the anchor, the registered Stage-R target) is determined by the anchor's logits, and therefore by the anchor's full next-token distribution p — the quantity D-5 names as the mandatory free baseline ("steering from the target's distribution at the last verified position, which every speculative decoder computes anyway"). sideband/FINDINGS.md:267 · d147e8c0cd6b

In plain words
CLOSED BY DATED ERRATUM
the lane was closed by a dated erratum, not by a new run
TARGET-FREE
the reconstruction target was itself free information
STANDS
the verdict is unchanged by later entries
FREE-ACTUATOR WIN
a steering arm fed only free information beat the plain speculative decoder
NULL
no effect found against the bar
PASS
the registered gate passed
VOID
void as a test: its premise no longer holds
STAGE-R PASS
the reconstruction stage passed its registered gate
index: stands
the registered verdict is unchanged and no later entry narrows it
Regime
MATHEMATICAL IDENTITY, numerically checked on REAL byproduct rows (Llama-3.1-8B-Instruct lm_head, untied and bias-free, read from the cached bf16 checkpoint as float32; 48 pinned rows = the first 4 blocks of the first 3 prompts of each of the four registered vanilla gamma-4 verdict runs, every npz sha-bound to its committed summary; CPU compute on the GPU box — … sideband/FINDINGS.md:267 · d147e8c0cd6b
The full regime line

MATHEMATICAL IDENTITY, numerically checked on REAL byproduct rows (Llama-3.1-8B-Instruct lm_head, untied and bias-free, read from the cached bf16 checkpoint as float32; 48 pinned rows = the first 4 blocks of the first 3 prompts of each of the four registered vanilla gamma-4 verdict runs, every npz sha-bound to its committed summary; CPU compute on the GPU box — the tool needs the cached weights and the gitignored raws; CI imports the module through tests/test_ha_identity.py, where torch and safetensors load lazily, and never runs it; the identity is generic to any linear unembedding of full column rank with vocab > d, tied or untied, and is measured here on one model). CLAIM: h_a (the final-norm output at the anchor, the registered Stage-R target) is determined by the anchor's logits, and therefore by the anchor's full next-token distribution p — the quantity D-5 names as the mandatory free baseline ("steering from the target's distribution at the last verified position, which every speculative decoder computes anyway"). W_U @ h reproduces the logged argmax 48/48 and the logged top-32 ids at 0.9948 mean overlap (min 0.969; max top-1 probability gap 0.0206 against the values logged by the live bf16 lm_head of the NF4-loaded verifier); the least-squares inverse from the float32 logits recovers h_a at median relative error 2.5e-8 (max 3.2e-8), the float32 rounding floor (from float64 logits 1.9e-14); closed-form least-squares recovery from only the top-k (id, log-prob) pairs of the recomputed logits (not the logged float16 pairs) reaches median per-row R^2 +1.0000 at k = 4097, +0.506 at 2048, +0.216 at 512, +0.078 at 128, +0.014 at 32 (median cosines 1.000 / 0.711 / 0.465 / 0.280 / 0.123) — a per-row R^2 about the row's own mean in raw h space from a prior-free solve, NOT Stage-R's standardized pooled-MSE R^2 and not comparable to the ridge arms' +0.258 / +0.324; the top-k recovery uses log-probabilities only with the log-partition as an extra unknown, so p alone, not the logits, determines h_a. CONDITIONING: W_U has full column rank (singular values 1.453 .. 94.5, condition 65); the all-ones vector, the softmax's free direction, lies 99.2% inside col(W_U), yet the augmented system [W_U, -1] has smallest singular value 1.451, so nothing is weakly determined; from bfloat16-rounded full log-probabilities with logZ unknown, h_a recovers to median relative error 3.5e-03 (max 6.9e-03). WHAT STANDS: every Stage-R and DR-1..DR-4 readout as pinned (F-22) — the gates were run against their registered arms and recomputed bit-exact. WHAT FALLS: their meaning for the lane. The registered FREE arm (657 anchor summaries; R^2 +0.083) and the DR-2/DR-4 token encodings (FREETOK +0.258, FREETOK2 +0.255) were lossy proxies for the free information D-5 names, from which the target follows exactly (R^2 1.000); the stream's reconstruction (TEL +0.253; UNION +0.324) is therefore K-2's "lossy copy of information the decoder already has", with no hashing caveat and no further comparator needed — "beyond p" is priced at this target: nothing is. The Stage-A telemetry arms (F = FREETOK reconstruction, T = UNION reconstruction) are struck pre-GPU: a host that computes p for rejection sampling can inject h_a itself with at most one least-squares solve against the cached lm_head (a 4096 x 128256 matvec plus a precomputed 4096 x 4096 inverse; this harness holds h_a directly through its final-norm catch). "Free" is D-5's convention; D-1's premise never priced shipping p or h_a to a disaggregated drafter (K-6 territory, not settled here). The oracle arm O (true h_a) is thereby the DEPLOYABLE free-actuator arm — SD^2-lite from the verifier's own distribution — whose V-vs-O readout sits outside the K-3 kill text and this lane's claim; it needs its own registration if run (label FREE-ACTUATOR WIN / NULL; a WIN is an SD^2 configuration result, not a sideband result). WHAT IT DOES NOT TOUCH, AND WHY: the existence results F-15/F-17 concern future-position concepts at offsets 2-3, which the anchor's p does not determine; but their free arm (a) is the same 657-summary proxy, so "beyond p?" arises there too, and the record's answer is the OQ-6 h-probe row of the same grids (margins_h_vs_a in results/real/probes_s1{,_replication}.json and probes_s3{,_replication}.json): under the identity a linear probe on h_a is a linear probe on the full logit vector, the strongest linear free-p baseline, and at the pinned L1 offset-2 cells it adds nothing over (a) — h-a margin S1 +0.0009 [-0.0087, +0.0107] primary, -0.0012 [-0.0112, +0.0090] replication; S3 -0.0026 [-0.0071, +0.0021], -0.0042 [-0.0093, +0.0015] — against c-a margins +0.0163 / +0.0141 (S1) and +0.0147 / +0.0173 (S3), and sits below (a) at offsets 3-4 on all four grids (offset-3 h-a -0.0205 / -0.0399 / -0.0111 / -0.0173). The K-2 margins are therefore not linear-logit information and F-15/F-17 stand; a NONLINEAR-in-p free baseline at the registered mlp256 capacity was never run and is the unrun K-2 comparator (not authorized here). The topic controls F-16/F-18 and the conversion nulls F-19/F-21 stand on their own. The only reconstruction target where telemetry could carry non-free information is h at FUTURE anchors (OQ-16's own anticipatory variant, never run); the fold-back-#2 hard stop forbids a third arc without the user's explicit exception. CLASSIFICATION: no partition branch is claimed. The limbo clause's condition is met (a Stage A never run after a PASS) and its rule, "neither limbo may ever be re-labeled a null", is honored; the catch-all is not invoked because the branch it would name (RECONSTRUCTION-CEILING NULL, "Stage-R leg 1 fails") is contradicted by the measured leg-1 pass; the disposition is the partition's own VOID precedent (no branch claimed, a dated note, the lane stops in practice under the fold-back-#2 hard stop). The lane is CLOSED BY DATED ERRATUM, mechanism TARGET-FREE; STAGE-R PASS stands as the mechanical outcome of its gate; the DR-3 OUTCOME's "reopens the Stage-A telemetry question" sentence is superseded. The lane ENDS. DISCLOSURE: the target was chosen for logging convenience (model.norm, one hook) rather than SD^2's mid-layer read (layers 3, L/2, L-2), which no identity of this kind ties to p (many-to-one; how much of those states p predicts is unmeasured); the free arm was built from the existing probe_input summaries rather than from D-5's definition; three diagnostics priced "beyond FREETOK's encoding" while the closed-form free comparator was one linear solve away. The observation came from the 2026-09-05 creative-lens adversarial pass over the gathered data (contrarian lens), reproduced by that pass's verifier, by an operator scratch check and by the committed tool; CRITIC RECOMPUTE ROUND (2026-09-05, before this commit): an independent implementation (float64 QR inverse; SVD least squares for the top-k solves; log-probabilities only) reproduced the forward numbers exactly, the full-logit inverse to 8 significant figures and every k < 4097 number to 7 or more digits; it exposed the first draft's k = 4097 row (median +0.9999, min 0.9988) as an artifact of an ill-conditioned normal-equations solve with a 1e-8 ridge, replaced by the committed SVD solve, and confirmed full column rank and the augmented-system conditioning above; it also blocked three sentences of the first draft — a "mechanical" catch-all classification against the limbo clause, an unargued "does not touch F-15/F-17", and a "standardized R^2" label — all rewritten before the commit. ATTESTATION: results/real/oq16_ha_identity.json carries git_commit 8542ef5-dirty; at run time the tracked dirt was this commit's own documentation and registration text, the tool and its test were untracked, and every input is sha-bound (lm_head shard sha256 and the twelve npz sha256s in the file).

Source
results/real/oq16_ha_identity.json, tools/ha_identity_check.py, docs/BLUEPRINT.md (OQ-16 DATED ERRATUM + LANE CLOSURE; STAGE-A FREEZE; …
The full source line

results/real/oq16_ha_identity.json, tools/ha_identity_check.py, docs/BLUEPRINT.md (OQ-16 DATED ERRATUM + LANE CLOSURE; STAGE-A FREEZE; DR-3 OUTCOME), DECISIONS.md (D-2, D-5), results/real/probes_s1.json, results/real/probes_s1_replication.json, results/real/probes_s3.json, results/real/probes_s3_replication.json (margins_h_vs_a), results/real/oq16_stage_r_dr3.json, results/real/oq16_stage_r_dr4.json, results/real/k3_verdict_arm1.summary.json, results/real/k3_verdict_s1_arm1.summary.json, results/real/oq15c_verdict_arm1.summary.json, results/real/oq15c_verdict_s3_arm1.summary.json (byproduct sha bindings; raws gitignored). Commit chain 2e44105 -> 786ab59 -> 8542ef5 -> this commit.

sideband/FINDINGS.md:268 · d147e8c0cd6b
Date
2026-09-05 sideband/FINDINGS.md:266 · d147e8c0cd6b
Index headline
argmax match 48/48; top-32 overlap 0.9948 mean; full-logit inverse relative error 2.5e-8 median; top-k closed-form per-row R^2 +1.0000 (k = 4097) / +0.506 (2048) / +0.216 (512) / +0.078 (128) / +0.014 (32)
48/48W_U @ h reproduces the logged argmax 48/48 and the logged top-32sideband/FINDINGS.md:267 · d147e8c0cd6b
0.9948argmax 48/48 and the logged top-32 ids at 0.9948 mean overlap (minsideband/FINDINGS.md:267 · d147e8c0cd6b
0.969top-32 ids at 0.9948 mean overlap (min 0.969; max top-1 probabilitysideband/FINDINGS.md:267 · d147e8c0cd6b
0.0206(min 0.969; max top-1 probability gap 0.0206 against the valuessideband/FINDINGS.md:267 · d147e8c0cd6b
2.5e-8recovers h_a at median relative error 2.5e-8 (max 3.2e-8), thesideband/FINDINGS.md:267 · d147e8c0cd6b
3.2e-8h_a at median relative error 2.5e-8 (max 3.2e-8), the float32 roundingsideband/FINDINGS.md:267 · d147e8c0cd6b
1.9e-14float32 rounding floor (from float64 logits 1.9e-14); closed-formsideband/FINDINGS.md:267 · d147e8c0cd6b
+1.0000float16 pairs) reaches median per-row R^2 +1.0000 at k = 4097, +0.506 atsideband/FINDINGS.md:267 · d147e8c0cd6b
sideband/FINDINGS.md:265 · d147e8c0cd6b sideband/docs/FINDINGS_INDEX.md:124 · aa88354d13cb
F-24OQ-17 FULL-FREE EXISTENCE RE-TESTsidebandindex: standsABSORBEDTWIN-ROBUSTOFF3-ABSORBEDCORRECTEDSURVIVES

The F-15 and F-17 existence margins are ABSORBED by the full free baseline at the registered reader on both harvests of both sources: no MATERIAL (+0.005) equal-dims margin; …

The full index summary

The F-15 and F-17 existence margins are ABSORBED by the full free baseline at the registered reader on both harvests of both sources: no MATERIAL (+0.005) equal-dims margin; telemetry on top of the full free baseline still adds about +0.003 (Mt) CI-clear on three of four harvests, matched by the corrected twin's own gain (Mn_r +0.0032 / +0.0020 at S1), under which Mt stops being evidence of telemetry information; the corrected-twin diagnostic read TWIN-ROBUST at both sources; the pinned prediction (S1 SURVIVES-FULL-FREE; S3 SPLIT or NOT-SHOWN) failed on both sources

Heading, as written OQ-17 FULL-FREE EXISTENCE RE-TEST: at the registered reader the K-2 existence margins of F-15 (S1) and F-17 (S3) are ABSORBED by the full free baseline — (a) ++ hashed token identities ++ z-scored h_a read at mlp256 — at the pinned L1 offset-2 cell on both harvests of both sources (every equal-dims D CI95 upper below +0.005; Mt beside it +0.003 or less); the pinned prediction failed on both sources; the registered equal-dims twin was found non-inert after the fact and a registered post-hoc corrected-twin diagnostic read TWIN-ROBUST at S1 and TWIN-ROBUST at S3

MECHANICAL LABELS: S1 ABSORBED (horizon OFF3-ABSORBED); S3 ABSORBED — both at the registered reader at the registered budget, both preconditions met on every harvest (reader validity: min gate-cell accuracy minus floor +0.0468 / +0.0393 at S1, +0.0272 / +0.0207 at S3; Mf gate CI upper +0.0475 / +0.0432 / +0.0298 / +0.0241; leak rider: … sideband/FINDINGS.md:275 · d147e8c0cd6b

The full first sentence

MECHANICAL LABELS: S1 ABSORBED (horizon OFF3-ABSORBED); S3 ABSORBED — both at the registered reader at the registered budget, both preconditions met on every harvest (reader validity: min gate-cell accuracy minus floor +0.0468 / +0.0393 at S1, +0.0272 / +0.0207 at S3; Mf gate CI upper +0.0475 / +0.0432 / +0.0298 / +0.0241; leak rider: shuffled fplus 0.5733 / 0.5681 / 0.5792 / 0.5851 against committed shuffled (a) 0.5889 / 0.5871 / 0.5799 / 0.5858, shuffled fplus_b 0.5740 / 0.5706 / 0.5787 / 0.5845, both legs pass on all four).

In plain words
ABSORBED
no material margin over the full free baseline at the registered reader, and no more than a scrambled block of the same size adds; the gate fired on neither harvest of the source
TWIN-ROBUST
the reading does not depend on which form of the noise control is used
OFF3-ABSORBED
the same absorbed reading three positions ahead instead of two
CORRECTED
a later dated erratum corrected a number or a sentence in this entry
SURVIVES
the margin cleared its floor under the registered rule
NON-CRITERION
a diagnostic that reports and prices but cannot change a verdict
WITHDRAWN
a sentence was withdrawn by a dated note; the text itself stays as written
FAILED
a pinned prediction failed
SURVIVES-FULL-FREE
the margin cleared the floor against the full free baseline at that cell
SPLIT
the gate fired on one of the source's two harvests and not on the other
NOT-SHOWN
the test could not show the effect; a non-result, not a negative result
index: stands
the registered verdict is unchanged and no later entry narrows it
Regime
REAL MODELS (probe training only on the four committed Phase-3 harvests — S1 primary seeds 0..7 and replication seeds 8..11 of the NF4 Llama-3.1-8B-Instruct shadow-tap harvest; S3 primary and replication of the OLMoE-1B-7B router harvest; no model forward, no acceptance, no wall clock). sideband/FINDINGS.md:273 · d147e8c0cd6b
The full regime line

REAL MODELS (probe training only on the four committed Phase-3 harvests — S1 primary seeds 0..7 and replication seeds 8..11 of the NF4 Llama-3.1-8B-Instruct shadow-tap harvest; S3 primary and replication of the OLMoE-1B-7B router harvest; no model forward, no acceptance, no wall clock). Registered 2026-09-06 in docs/BLUEPRINT.md "OQ-17 FULL-FREE EXISTENCE RE-TEST REGISTRATION" before any registered cell ran, critic-gated twice pre-freeze (commit 1d1a7f8), CRITERION-BEARING for F-15 / F-17 through pinned consequence rules. Arms per harvest, identical block order: fplus = (a) 657 logged summaries of p_a ++ TOK (multi-hot of the anchor's top-32 next-token ids hashed mod 2048, the F-22 DR-2 rule) ++ hz (h_a, the final-norm state that produced p_a, clipped to the TRAIN support and z-scored per channel with TRAIN-split moments: S1 4096 dims, 1200 / 600 train rows, 304,461 / 152,763 positions, per-channel sd 0.97..10.29 median 2.11, support to 56.25; S3 2048 dims, 295,346 / 148,876 positions, sd 0.67..6.65 median 2.05, support to 51.25; moments sha256 in every payload); fplus_b = fplus ++ aligned telemetry (S1 2048 slot dims, S3 1024 expert dims); fplus_n = fplus ++ the same row's telemetry at anchor (a+17) mod row-length (the F-16 shift twin; equal dims to fplus_b by construction: S1 8849, S3 5777; fplus 6801 / 4753). Trainer, budget, estimand and bootstrap byte-identical to Phase 3 (30 epochs, VAL-selected epoch; per-prompt paired successes summed over train seeds {0,1,2}, ONE prompt-cluster bootstrap CI95 over the 30 test prompts, n_boot 2000, seed 0). Cells: L1 x off2 x {linear, mlp256} x 3 seeds x 3 arms on every harvest (18), plus L1 x off3 x mlp256 on the two S1 harvests (9; offsets bound to the source tag); a two-leg leak rider first (shuffled fplus vs the committed shuffled (a); shuffled fplus_b vs shuffled fplus; band 0.01). READER-VALIDITY precondition: all nine gate-cell probes hold the majority floor and Mf's gate CI upper >= 0. Gate: L1 / off2 / mlp256 D = acc(fplus_b) - acc(fplus_n) under the Phase-3 triple rule (CI95 lower > 0, all three seeds positive, point >= +0.005); labels per source from BOTH harvests (tools/run_fullfree.py --conclude-from, mechanical): READER-INVALID / SURVIVES-FULL-FREE / SPLIT / ABSORBED (fires on neither AND every D row's CI95 upper < +0.005 on both harvests) / NOT-SHOWN / INCOMPLETE, all scoped to the registered reader at the registered budget; Mt = acc(fplus_b) - acc(fplus) quoted beside D everywhere; Mn = acc(fplus_n) - acc(fplus), Mf = acc(fplus) - acc(a) and Mc = acc(fplus) - acc(c) reported against the committed Phase-3 cells (paired through the sha-bound state files; the committed payload margins re-derived at load to 1e-12). Post-hoc, NON-CRITERION, registered before any of its registered cells ran (same block, "OQ-17 POST-HOC TWIN-INERTNESS DIAGNOSTIC", critic-gated pre-run, committed 675a8f6): arm fplus_r = fplus ++ the slots of the same-generation-index row of a DIFFERENT prompt of the same split at the same anchor index (a prompt-level within-split derangement per train seed); the gate cell only, three seeds, all four harvests; readouts D_r = acc(fplus_b) - acc(fplus_r), Mn_r = acc(fplus_r) - acc(fplus), T = acc(fplus_n) - acc(fplus_r); readings TWIN-ROBUST / TWIN-MASKED (clean fires only) / TWIN-SPLIT / TWIN-DEPENDENT / TWIN-INCOMPLETE plus a T reading; it cannot change a registered label — it pins which qualifier the consequence text carries. Provenance: the four registered payloads bind commit 1d1a7f8 with resumed_cells 0 and partial false; the S3 REPLICATION payload records 1d1a7f8-dirty with git_status_tracked = [.gitignore, OPEN_QUESTIONS.md, docs/BLUEPRINT.md] — the twin-diagnostic registration written while that run was in flight; tools/, src/ and tests/ unchanged against HEAD (critic-verified by git diff at write-up; the instrument module was imported at process start, before the edits; no artifact records the check — critic-attested, the F-17 operator-attested class). The four twin payloads bind 675a8f6 and record the instrument's sha256 and the per-seed permutation sha256. Critic rounds, recompute mode, all before this text: S1 pair and S3 pair (every payload number reproduced from the state files to 0.0 by the repo bootstrap and an independent numpy re-implementation; harvest npz shas, floors, split sizes, train-only moments and the committed Phase-3 pairings re-derived), a pre-run design review of the twin diagnostic (four blockers applied before launch), and a recompute of the four twin payloads.

Source
results/real/fullfree_s1.json, fullfree_s1.json.state.jsonl, fullfree_s1_r.json, fullfree_s1_r.json.state.jsonl, fullfree_s3.json, fullfree_s3.json.state.jsonl, fullfree_s3_r.json, fullfree_s3_r.json.state.jsonl, fullfree_s1_twin.json, fullfree_s1_twin.json.state.jsonl, fullfree_s1_r_twin.json, …
The full source line

results/real/fullfree_s1.json, fullfree_s1.json.state.jsonl, fullfree_s1_r.json, fullfree_s1_r.json.state.jsonl, fullfree_s3.json, fullfree_s3.json.state.jsonl, fullfree_s3_r.json, fullfree_s3_r.json.state.jsonl, fullfree_s1_twin.json, fullfree_s1_twin.json.state.jsonl, fullfree_s1_r_twin.json, fullfree_s1_r_twin.json.state.jsonl, fullfree_s3_twin.json, fullfree_s3_twin.json.state.jsonl, fullfree_s3_r_twin.json, fullfree_s3_r_twin.json.state.jsonl, tools/run_fullfree.py, tools/run_fullfree_twin.py, tests/test_fullfree.py, tests/test_fullfree_twin.py, docs/BLUEPRINT.md (OQ-17 registration, twin diagnostic, OQ-17 OUTCOME), results/real/probes_s1.json.state.jsonl, probes_s1_replication.json.state.jsonl, probes_s3.json.state.jsonl, probes_s3_replication.json.state.jsonl (the committed (a)/(c) cells paired against). Commit chain 1d1a7f8 -> 675a8f6 -> this commit.

sideband/FINDINGS.md:274 · d147e8c0cd6b
Date
2026-09-06 sideband/FINDINGS.md:272 · d147e8c0cd6b
Index headline
gate D S1 +0.0002 [-0.0012, +0.0015] / +0.0012 [-0.0003, +0.0028]; S3 +0.0003 [-0.0013, +0.0020] / +0.0001 [-0.0019, +0.0021]; every D CI95 upper below +0.005; …
The full headline

gate D S1 +0.0002 [-0.0012, +0.0015] / +0.0012 [-0.0003, +0.0028]; S3 +0.0003 [-0.0013, +0.0020] / +0.0001 [-0.0019, +0.0021]; every D CI95 upper below +0.005; Mt S1 +0.0034 [+0.0021, +0.0048] / +0.0033 [+0.0012, +0.0052], S3 +0.0033 [+0.0013, +0.0054] / +0.0013 [-0.0008, +0.0035]

+0.005(every equal-dims D CI95 upper below +0.005; Mt beside it +0.003 orsideband/FINDINGS.md:271 · d147e8c0cd6b
+0.003D CI95 upper below +0.005; Mt beside it +0.003 or less); the pinnedsideband/FINDINGS.md:271 · d147e8c0cd6b
+0.0468min gate-cell accuracy minus floor +0.0468 / +0.0393 at S1,sideband/FINDINGS.md:275 · d147e8c0cd6b
+0.0393gate-cell accuracy minus floor +0.0468 / +0.0393 at S1, +0.0272 /sideband/FINDINGS.md:275 · d147e8c0cd6b
+0.0272minus floor +0.0468 / +0.0393 at S1, +0.0272 / +0.0207 at S3; Mfsideband/FINDINGS.md:275 · d147e8c0cd6b
+0.0207floor +0.0468 / +0.0393 at S1, +0.0272 / +0.0207 at S3; Mf gate CI uppersideband/FINDINGS.md:275 · d147e8c0cd6b
+0.0475+0.0272 / +0.0207 at S3; Mf gate CI upper +0.0475 / +0.0432 / +0.0298 /sideband/FINDINGS.md:275 · d147e8c0cd6b
+0.0432/ +0.0207 at S3; Mf gate CI upper +0.0475 / +0.0432 / +0.0298 / +0.0241;sideband/FINDINGS.md:275 · d147e8c0cd6b
sideband/FINDINGS.md:271 · d147e8c0cd6b sideband/docs/FINDINGS_INDEX.md:125 · aa88354d13cb
F-25OQ-18 COARSE-LABEL FULL-FREE RE-TESTsidebandindex: standsABSORBEDNOT-SHOWNSPLITTWIN-MATCHEDOFF3-ABSORBED

At the registered reader with the corrected twin: S1 L2 ABSORBED, S1 L3 NOT-SHOWN, S3 L2 ABSORBED, S3 L3 SPLIT (fired on the replication harvest only; the entry's post-hoc line reads the fire as inflated by up to 0.0067 by the twin's own cost, not as half a survival); prediction 3/4 held on labels; the L1 verdict (F-24) untouched; …

The full index summary

At the registered reader with the corrected twin: S1 L2 ABSORBED, S1 L3 NOT-SHOWN, S3 L2 ABSORBED, S3 L3 SPLIT (fired on the replication harvest only; the entry's post-hoc line reads the fire as inflated by up to 0.0067 by the twin's own cost, not as half a survival); prediction 3/4 held on labels; the L1 verdict (F-24) untouched; no all-space sentence issues

Heading, as written OQ-18 COARSE-LABEL FULL-FREE RE-TEST: at the registered reader the F-15 / F-17 margins at the coarse label spaces read S1 L2 ABSORBED; S1 L3 NOT-SHOWN; S3 L2 ABSORBED; S3 L3 SPLIT (every L3 label read at the L3 group lens (S1: one mega-group of 455/510 concepts including all 165 zero-degree concepts, the next group 10; S3: two mega-groups of 268 and 193 = 461/510 including all 20 zero-degree concepts)) against the full free baseline ((a) ++ hashed token ids ++ z-scored h_a) with the corrected twin, on both harvests of both sources

[read under the 2026-09-07 rider amendment after the registered rider fired at S1 L3 (raw leg 2 +0.0132 against the 0.01 band); diagnostic branch AMENDABLE, mechanism TWIN-MATCHED; a post-hoc reroute around a fired kill condition: the leg that fired was replaced after the fire by a twin-referenced dependence-excess leg whose S1 seed-0 values at both spaces were known from the diagnostic payload when the amendment was written; it is a fresh test only at the other three harvests; … sideband/FINDINGS.md:286 · d147e8c0cd6b

The full first sentence

[read under the 2026-09-07 rider amendment after the registered rider fired at S1 L3 (raw leg 2 +0.0132 against the 0.01 band); diagnostic branch AMENDABLE, mechanism TWIN-MATCHED; a post-hoc reroute around a fired kill condition: the leg that fired was replaced after the fire by a twin-referenced dependence-excess leg whose S1 seed-0 values at both spaces were known from the diagnostic payload when the amendment was written; it is a fresh test only at the other three harvests; the E leg sees every non-causal block the raw leg sees (the twin lacks it) and is sharper under a drifted reader; what it drops (prediction-marginal drift and twin-shared position information) are non-leaks; S1 diagnostic X_raw at seed 0 +0.0006, the registered statistic, twin-referenced, would also have passed] MECHANICAL LABELS per (source, space), both harvests, registered reader at the registered budget: S1 L2 ABSORBED; S1 L3 NOT-SHOWN; S3 L2 ABSORBED; S3 L3 SPLIT (every L3 label read at the L3 group lens (S1: one mega-group of 455/510 concepts including all 165 zero-degree concepts, the next group 10; S3: two mega-groups of 268 and 193 = 461/510 including all 20 zero-degree concepts)).

In plain words
ABSORBED
no material margin over the full free baseline at the registered reader, and no more than a scrambled block of the same size adds; the gate fired on neither harvest of the source
NOT-SHOWN
the test could not show the effect; a non-result, not a negative result
SPLIT
the gate fired on one of the source's two harvests and not on the other
OFF3-ABSORBED
the same absorbed reading three positions ahead instead of two
NON-CRITERION
a diagnostic that reports and prices but cannot change a verdict
OFF3-NOT-SHOWN
three positions ahead the test could not show the effect
OFF3-SURVIVES
the margin cleared its floor three positions ahead
SURVIVES-FULL-FREE
the margin cleared the floor against the full free baseline at that cell
FAILED
a pinned prediction failed
SURVIVES
the margin cleared its floor under the registered rule
OFF4-ABSORBED
the same absorbed reading four positions ahead
PASS
the registered gate passed
index: stands
the registered verdict is unchanged and no later entry narrows it
Regime
REAL MODELS (probe training only on the four committed Phase-3 harvests — S1 primary seeds 0..7 and replication seeds 8..11 of the NF4 Llama-3.1-8B-Instruct shadow-tap harvest; S3 primary and replication of the OLMoE-1B-7B router harvest; no model forward, no acceptance, no wall clock). sideband/FINDINGS.md:284 · d147e8c0cd6b
The full regime line

REAL MODELS (probe training only on the four committed Phase-3 harvests — S1 primary seeds 0..7 and replication seeds 8..11 of the NF4 Llama-3.1-8B-Instruct shadow-tap harvest; S3 primary and replication of the OLMoE-1B-7B router harvest; no model forward, no acceptance, no wall clock). Registered 2026-09-06 in docs/BLUEPRINT.md "OQ-18 COARSE-LABEL FULL-FREE RE-TEST REGISTRATION" before any registered cell ran, critic-gated pre-freeze (freeze commit 15c4e9c), CRITERION-BEARING for F-15 / F-17 at their L2 / L3 cells through the pinned consequence rules. Arms per harvest, identical block order: fplus = (a) 657 logged summaries of p_a ++ TOK (hashed top-32 next-token ids) ++ hz (h_a clipped to the train support, z-scored with train moments); fplus_b = fplus ++ aligned telemetry; fplus_r = fplus ++ the CORRECTED prompt-level-deranged twin (label-inert by construction, not accuracy-inert); fresh (a) cells at mlp256; the committed linear (a) / (c) cells paired at linear. Label spaces L2 (k512 concepts + merged glue) and L3 (transition-coherent groups; S1 one mega-group of 455/510 with all 165 zero-degree concepts, S3 two mega-groups of 268 and 193 = 461/510 with all 20); offsets bound per space to the F-15 / F-17 firing sets (S1 L2 {2,3}, S1 L3 {2,3}, S3 L2 {2,3}, S3 L3 {2,3,4}); off2 at linear and mlp256, further offsets at mlp256; train seeds {0,1,2}; the registered Phase-3 trainer and budget (30 epochs, VAL-selected epoch); per-space leak rider with a fresh shuffled (a) reference; the three-leg reader-validity precondition (twelve probes on the floor, Mf vs the fresh (a), Mf_x vs the committed linear (a)); gate off2 / mlp256 D = acc(fplus_b) - acc(fplus_r) under the Phase-3 triple rule (CI95 lower > 0, all seeds positive, point >= +0.005), labels per (source, space) from both harvests; Mt and Mn_r quoted beside D; estimand and bootstrap as Phase 3 (seed-summed per-prompt successes, paired_cluster_delta_ci, n_boot 2000, seed 0, 30 test prompts).

Source
results/real/fullfree_coarse_s1.json, fullfree_coarse_s1.json.state.jsonl, fullfree_coarse_s1_r.json, fullfree_coarse_s1_r.json.state.jsonl, fullfree_coarse_s3.json, fullfree_coarse_s3.json.state.jsonl, fullfree_coarse_s3_r.json, fullfree_coarse_s3_r.json.state.jsonl (sha-header bound to their harvests; committed with this finding); …
The full source line

results/real/fullfree_coarse_s1.json, fullfree_coarse_s1.json.state.jsonl, fullfree_coarse_s1_r.json, fullfree_coarse_s1_r.json.state.jsonl, fullfree_coarse_s3.json, fullfree_coarse_s3.json.state.jsonl, fullfree_coarse_s3_r.json, fullfree_coarse_s3_r.json.state.jsonl (sha-header bound to their harvests; committed with this finding); tools/run_fullfree_coarse_amended.py (the amended instrument, sha aafd3ae84998..., importing tools/run_fullfree_coarse.py byte-identical at its frozen sha) and tests/test_fullfree_coarse_amended.py (the committed amended payloads reproduce their own conclusion, resume plan, attestation and v1 preservation; the frozen test file skips amended payloads by a dated note); results/real/fullfree_coarse_s3.rider_v1.json, fullfree_coarse_s3_r.rider_v1.json (frozen-instrument payloads, preserved, never concluded); results/real/oq18_rider_diag_s1.json + state file and docs/OQ18_RIDER_DIAGNOSTIC.md (the fire's diagnostic); results/real/oq18_rider_poscontrol_s1.json + state file and docs/OQ18_RIDER_POSCONTROL.md (the positive control); docs/BLUEPRINT.md dated amendment bullet (2026-09-07 / 2026-09-13); docs/BLUEPRINT.md OQ-18 registration and OUTCOME blocks; the committed Phase-3 state files for the paired linear cells (results/real/probes_s1{,_replication}.json.state.jsonl, probes_s3{,_replication}.json.state.jsonl).

sideband/FINDINGS.md:285 · d147e8c0cd6b
Date
2026-09-14 sideband/FINDINGS.md:281 · d147e8c0cd6b
Index headline
gate D S1 L2 +0.0011 [-0.0001, +0.0023] / +0.0006 [-0.0007, +0.0020]; S1 L3 +0.0010 [-0.0006, +0.0026] / +0.0012 [-0.0015, +0.0040]; S3 L2 +0.0018 [+0.0008, +0.0030] / +0.0003 [-0.0009, +0.0013]; …
The full headline

gate D S1 L2 +0.0011 [-0.0001, +0.0023] / +0.0006 [-0.0007, +0.0020]; S1 L3 +0.0010 [-0.0006, +0.0026] / +0.0012 [-0.0015, +0.0040]; S3 L2 +0.0018 [+0.0008, +0.0030] / +0.0003 [-0.0009, +0.0013]; S3 L3 +0.0041 [+0.0026, +0.0055] / +0.0053 [+0.0019, +0.0091] with Mt +0.0010 [-0.0012, +0.0030] / -0.0014 [-0.0041, +0.0016]

455/510at the L3 group lens (S1: one mega-group of 455/510 concepts including allsideband/FINDINGS.md:280 · d147e8c0cd6b
461/51010; S3: two mega-groups of 268 and 193 = 461/510 including all 20sideband/FINDINGS.md:280 · d147e8c0cd6b
+0.0132registered rider fired at S1 L3 (raw leg 2 +0.0132 against the 0.01 band);sideband/FINDINGS.md:286 · d147e8c0cd6b
0.01at S1 L3 (raw leg 2 +0.0132 against the 0.01 band); diagnosticsideband/FINDINGS.md:286 · d147e8c0cd6b
+0.0006non-leaks; S1 diagnostic X_raw at seed 0 +0.0006, the registeredsideband/FINDINGS.md:286 · d147e8c0cd6b
455/510at the L3 group lens (S1: one mega-group of 455/510 concepts including allsideband/FINDINGS.md:286 · d147e8c0cd6b
461/51010; S3: two mega-groups of 268 and 193 = 461/510 including all 20sideband/FINDINGS.md:286 · d147e8c0cd6b
+0.0011 [-0.0001, +0.0023]concepts)). S1 L2: ABSORBED. Gate D +0.0011 [-0.0001, +0.0023] / +0.0006 [-0.0007,sideband/FINDINGS.md:286 · d147e8c0cd6b
sideband/FINDINGS.md:280 · d147e8c0cd6b sideband/docs/FINDINGS_INDEX.md:126 · aa88354d13cb
F-26OQ-9 STAGE 1 (the prefetch halfsidebandindex: standsNON-CRITERIONPREFETCH-HEADROOMADDRESS-BEYOND-TOKENNOT-SHOWN

The router address stream is a prefetch source with headroom over a recency cache and a material margin over the free input token at B = 16, j = 1 (stage PREFETCH-HEADROOM, the seed replication agrees on every label); the margin is below the +0.05 floor at B = 32 at j = 1 and above it at j = 2 / 3 / 4 (+0.0545 / +0.0534 at j = 2; CI lower above the floor at j = 3 / 4 only); …

The full index summary

The router address stream is a prefetch source with headroom over a recency cache and a material margin over the free input token at B = 16, j = 1 (stage PREFETCH-HEADROOM, the seed replication agrees on every label); the margin is below the +0.05 floor at B = 32 at j = 1 and above it at j = 2 / 3 / 4 (+0.0545 / +0.0534 at j = 2; CI lower above the floor at j = 3 / 4 only); the critic's scope objection (the equally free previous token is not in the one-token baseline and gate 0 routes on it; F-26's GATE-0 NOTE records an unresolved alternative, that the stored gate-0 slot lags one forward) was answered by the equal-dims free-bigram twin on 2026-10-01: SURVIVES at that encoding, with a coverage caveat (the material part of the surviving margin sits on positions whose token key the train tables never saw); nothing reads it as context information or as support for the sideband premise

Heading, as written OQ-9 STAGE 1 (the prefetch half; NON-CRITERION, a K-6 datum): on the OLMoE S3 harvests the router address stream is a prefetch source with headroom over the recency cache and a material margin over the free INPUT TOKEN ALONE (a one-token train table; the equally free previous token is not in the baseline and gate 0 routes on it, so the equal-dims free-bigram twin is the unrun next rung) at B = 16 of 64 — stage label PREFETCH-HEADROOM on the primary harvest, the seed replication agrees on every label; ADDRESS-BEYOND-TOKEN +0.1032 / +0.1072 at B = 16 and below the +0.05 floor at B = 32; INTRA-FORWARD +0.4861 / +0.4816; DRAFT (perfect-draft ceiling, horizon 2) +0.1556 / +0.1545

MECHANICAL LABELS from the primary harvest's TEST rows at B = 16, the replication computed identically (primary / replication): STAGE LABEL PREFETCH-HEADROOM, replication AGREES. sideband/FINDINGS.md:292 · d147e8c0cd6b

In plain words
NON-CRITERION
a diagnostic that reports and prices but cannot change a verdict
PREFETCH-HEADROOM
a prefetch readout showed room above its free baseline at the registered cell
ADDRESS-BEYOND-TOKEN
the address history carried information beyond the current token at the registered cell
NOT-SHOWN
the test could not show the effect; a non-result, not a negative result
index: stands
the registered verdict is unchanged and no later entry narrows it
Regime
REAL MODELS (CPU-only readouts on the two committed S3 harvests of allenai/OLMoE-1B-7B-0125-Instruct, NF4 double-quant with bf16 routers, target-only sampled generations at temperature 1.0, 256 new tokens, the 210 committed short prompts; primary seeds 0..7 (results/real/harvest_moe), replication seeds 8..11 (harvest_moe_r) on the SAME prompts; … sideband/FINDINGS.md:290 · d147e8c0cd6b
The full regime line

REAL MODELS (CPU-only readouts on the two committed S3 harvests of allenai/OLMoE-1B-7B-0125-Instruct, NF4 double-quant with bf16 routers, target-only sampled generations at temperature 1.0, 256 new tokens, the 210 committed short prompts; primary seeds 0..7 (results/real/harvest_moe), replication seeds 8..11 (harvest_moe_r) on the SAME prompts; the sorted top-8 expert ids at the 16 router gates, no router weights; no model forward, no acceptance, no wall clock). Registered 2026-09-14 in docs/BLUEPRINT.md "OQ-9 STAGE-1 PREFETCH READOUTS REGISTRATION" before any readout ran on either harvest, critic-gated through three design passes and a final confirmation, frozen at c4c4511; NON-CRITERION for every existing finding (no K-2 / K-3 verdict, no F-15 / F-17 / F-24 / F-25 sentence and no existence or conversion label is touched); a systems datum for the K-6 memo. ESTIMAND: HIT RATE at residency budget B in {8, 16, 32} of the 64 experts per gate — the number of a target set's eight experts inside the resident set, summed over positions, divided by 8 times the positions; a COUNT, never bytes or latency; B = 16 is the label budget, 8 and 32 are reported. Readers fit on the train prompts (prompt_idx % 7 in {0..4}) pooled over seeds, lambda selected on val (5) from the widened grid {0.1, 1, 10, 100, 1000, 1e4, 1e5}, every number read on the 30 TEST prompts (6) pooled over seeds and positions with prompt-cluster bootstrap CI95 (n_boot 2000, seed 0); paired deltas on identical positions. ARMS per horizon j (S[t] the routing of the forward whose softmax is p_t; tokens[t] sampled from p_t is the FREE bonus token and feeds the forward whose routing is S[t+1] — the convention verified against the harvester pre-freeze and confirmed on the data by the critic's val-only diagnostic at landing): RECENCY_B = the B most recently used experts per gate (the cache policy); MARGINAL_B; FREE_TABLE = zero-fit top-B of TABLE(tokens[t]), the train token-conditional routing frequencies of the forward a token feeds (leave-one-prompt-out at train rows); FREE_TOK = ridge on TABLE(tokens[t]) — a ONE-TOKEN free arm, the D-5 analogue for prefetch as registered; HISTORY = ridge on the S[t] multi-hot, address-only; HISTORY_TOK = S[t] ++ TABLE(tokens[t]); at j = 2 only DRAFT_FREE / DRAFT_FULL add TABLE(tokens[t+1]), the TRUE next token as the D-2-named ceiling draft; INTRA = gate l from gates 0..l-k of the SAME forward (leads 1, 2, 4) against MARGINAL; (iv) a front-loading curve reported only. Readings SIGNAL iff paired CI95 lower > 0 and point >= +0.05 (the pinned materiality floor), NEGATIVE iff CI95 upper < 0, else NOT-SHOWN; HISTORY = HISTORY minus RECENCY_16 (j = 1), ADDRESS-BEYOND-TOKEN = HISTORY_TOK minus FREE_TOK (j = 1), INTRA-FORWARD = INTRA lead 1 minus MARGINAL_16 over target gates 1..15, DRAFT = DRAFT_FULL minus HISTORY_TOK (j = 2); CACHE-SUFFICIENT iff RECENCY_16 at j = 1 >= 0.90; the sideband sentence follows ADDRESS-BEYOND-TOKEN alone. Every label is scoped to the registered linear readers on the sorted top-8 id stream at hit rate at the registered budgets.

Source
results/real/prefetch_s3.json, prefetch_s3.json.state.jsonl, prefetch_s3.meta.json (primary, the VERDICT harvest); results/real/prefetch_s3_r.json, prefetch_s3_r.json.state.jsonl, prefetch_s3_r.meta.json (replication, confirmatory) — the state files sha-header bound to the shard shas recorded in the committed shard metas under results/real/harvest_moe (moe_s0..s7) and results/real/harvest_moe_r (moe_s8..s11); …
The full source line

results/real/prefetch_s3.json, prefetch_s3.json.state.jsonl, prefetch_s3.meta.json (primary, the VERDICT harvest); results/real/prefetch_s3_r.json, prefetch_s3_r.json.state.jsonl, prefetch_s3_r.meta.json (replication, confirmatory) — the state files sha-header bound to the shard shas recorded in the committed shard metas under results/real/harvest_moe (moe_s0..s7) and results/real/harvest_moe_r (moe_s8..s11); the .npz shards themselves are gitignored and tests/test_prefetch_landing.py checks the header against those metas; committed with this finding; tools/run_prefetch_readouts.py (the frozen instrument, sha 8529532d4d78..., numpy only, --conclude-from the exact recompute surface); tests/test_prefetch_readouts.py; tests/test_prefetch_landing.py (the committed payloads reproduce their own conclusion; the instrument on disk is the frozen one); docs/BLUEPRINT.md registration block with its PRE-FREEZE ROUND (four disclosed smokes on replication shard s8's val prompts, every one recomputed from its state file) and OUTCOME paragraph (the critic's landing recompute and its val-only alignment diagnostic are described there; the diagnostic's outputs are scratch, not evidence).

sideband/FINDINGS.md:291 · d147e8c0cd6b
Date
2026-09-15 sideband/FINDINGS.md:289 · d147e8c0cd6b
Index headline
B = 16, j = 1: recency 0.5197 / 0.5155, free one-token table 0.8111 / 0.8036, history 0.8064 / 0.8056, history ++ token 0.9143 / 0.9108; HISTORY +0.2867 [+0.2784, +0.2939] / +0.2901 [+0.2808, +0.2978]; ADDRESS-BEYOND-TOKEN +0.1032 [+0.0926, +0.1146] / +0.1072 [+0.0952, +0.1206] (B = 32, j = 1: +0.0405 / +0.0434; B = 32, j = 3 / 4: …
The full headline

B = 16, j = 1: recency 0.5197 / 0.5155, free one-token table 0.8111 / 0.8036, history 0.8064 / 0.8056, history ++ token 0.9143 / 0.9108; HISTORY +0.2867 [+0.2784, +0.2939] / +0.2901 [+0.2808, +0.2978]; ADDRESS-BEYOND-TOKEN +0.1032 [+0.0926, +0.1146] / +0.1072 [+0.0952, +0.1206] (B = 32, j = 1: +0.0405 / +0.0434; B = 32, j = 3 / 4: +0.0683 / +0.0654, +0.0698 / +0.0669); INTRA-FORWARD +0.4861 [+0.4710, +0.5009] / +0.4816 [+0.4663, +0.4975]; DRAFT ceiling (horizon 2) +0.1556 [+0.1512, +0.1599] / +0.1545 [+0.1495, +0.1594]; twin (2026-10-01): D_eq +0.0609 [+0.0514, +0.0712] / +0.0659 [+0.0547, +0.0788], D_add +0.0661 [+0.0569, +0.0762] / +0.0705 [+0.0592, +0.0834], at B = 32 (j = 1, the twin's only horizon; the j = 2..4 margins are over the one-token table and were never tested against the bigram twin) D_eq +0.0261 / +0.0287; covered-key D_eq +0.0454 / +0.0472 on val (unregistered diagnostic, not evidence); stratified read (2026-10-02, test split, B = 16, j = 1): covered D_eq +0.0463 [+0.0398, +0.0541] / +0.0481 [+0.0411, +0.0564], unseen-key +0.1940 / +0.1741; D_eq^PT +0.0532 [+0.0455, +0.0619] / +0.0565 [+0.0480, +0.0664], D_add^PT +0.0583 / +0.0609; covered positions by the smaller key's train occurrences one / two to four / five or more +0.0658 / +0.0552 / +0.0438 (replication +0.0683 / +0.0503 / +0.0456)

+0.1032agrees on every label; ADDRESS-BEYOND-TOKEN +0.1032 / +0.1072 at B = 16 andsideband/FINDINGS.md:288 · d147e8c0cd6b
+0.1072every label; ADDRESS-BEYOND-TOKEN +0.1032 / +0.1072 at B = 16 and below thesideband/FINDINGS.md:288 · d147e8c0cd6b
+0.05+0.1032 / +0.1072 at B = 16 and below the +0.05 floor at B = 32;sideband/FINDINGS.md:288 · d147e8c0cd6b
+0.4861the +0.05 floor at B = 32; INTRA-FORWARD +0.4861 / +0.4816; DRAFTsideband/FINDINGS.md:288 · d147e8c0cd6b
+0.4816floor at B = 32; INTRA-FORWARD +0.4861 / +0.4816; DRAFT (perfect-draftsideband/FINDINGS.md:288 · d147e8c0cd6b
+0.1556DRAFT (perfect-draft ceiling, horizon 2) +0.1556 / +0.1545sideband/FINDINGS.md:288 · d147e8c0cd6b
+0.1545ceiling, horizon 2) +0.1556 / +0.1545sideband/FINDINGS.md:288 · d147e8c0cd6b
0.5197 [0.5144, 0.5258]did not fire: RECENCY_16 at j = 1 0.5197 [0.5144, 0.5258] / 0.5155 [0.5092,sideband/FINDINGS.md:292 · d147e8c0cd6b
sideband/FINDINGS.md:288 · d147e8c0cd6b sideband/docs/FINDINGS_INDEX.md:127 · aa88354d13cb
F-27SEED-NOISE FLOOR (a methods datumsidebandindex: standsNON-CRITERIONFLOOR-MATERIALBAR-RESOLVED@210RULE-LEAKS@30NOTE-STANDS

a same-policy seed change moves pooled tokens/call by a CI95 half-width of 0.0385 / 0.0398 at 210 prompts and 0.0982 / 0.1032 at 30; seed noise alone clears +0.062 in 0.1060 / 0.1320 of 30-prompt subsets; …

The full index summary

a same-policy seed change moves pooled tokens/call by a CI95 half-width of 0.0385 / 0.0398 at 210 prompts and 0.0982 / 0.1032 at 30; seed noise alone clears +0.062 in 0.1060 / 0.1320 of 30-prompt subsets; a cross-seed, same-policy, single-comparator null (the record's legs are same-seed and all-comparator, that null unmeasured)

Heading, as written SEED-NOISE FLOOR (a methods datum; NON-CRITERION): on the four committed vanilla gamma-4 verdict runs a same-policy seed change moves pooled tokens/call by a CI95 half-width of 0.0385 / 0.0398 at 210 prompts and 0.0982 / 0.1032 at 30 (K3 / OQ15); one seed's per-prompt tokens/call correlates 0.4232 / 0.4172 with the other seed's; at n = 30 a +0.062 point win appears by seed noise alone in 0.1060 / 0.1320 of subsets and a single comparator leg of the F-19 WIN rule passes in 0.0235 / 0.0375 — FLOOR-MATERIAL on both populations, BAR-RESOLVED@210, RULE-LEAKS@30; a cross-seed, same-policy, single-comparator null

MECHANICAL LABELS (K3 the label population; OQ15 computed identically, the replication AGREES): FLOOR-MATERIAL on K3 (f_one at n = 30 0.1060, SE 0.0065, against the 0.02 kill threshold pinned from IDEAS rank 1) and FLOOR-MATERIAL on OQ15 (0.1320, SE 0.0070); STAGE NOTE-STANDS; readings BAR-RESOLVED@210 on both (the n = 210 half-width 0.0385 / 0.0398 against the 0.062 bar) and RULE-LEAKS@30 on both (g_one at n = 30 0.0235 / 0.0375, SE 0.0033 / 0.0042; the guarded rate is the same number by algebra, n_drafted = 4 n_blocks on every prompt); … sideband/FINDINGS.md:304 · d147e8c0cd6b

The full first sentence

MECHANICAL LABELS (K3 the label population; OQ15 computed identically, the replication AGREES): FLOOR-MATERIAL on K3 (f_one at n = 30 0.1060, SE 0.0065, against the 0.02 kill threshold pinned from IDEAS rank 1) and FLOOR-MATERIAL on OQ15 (0.1320, SE 0.0070); STAGE NOTE-STANDS; readings BAR-RESOLVED@210 on both (the n = 210 half-width 0.0385 / 0.0398 against the 0.062 bar) and RULE-LEAKS@30 on both (g_one at n = 30 0.0235 / 0.0375, SE 0.0033 / 0.0042; the guarded rate is the same number by algebra, n_drafted = 4 n_blocks on every prompt); neither label within two SEs of its threshold; the precondition held, nineteen of nineteen checks on both.

In plain words
NON-CRITERION
a diagnostic that reports and prices but cannot change a verdict
FLOOR-MATERIAL
seed noise alone can produce a win the size of the bar often enough to matter at small prompt counts
index: stands
the registered verdict is unchanged and no later entry narrows it
Regime
REAL MODELS (a CPU recompute on committed evidence, no model forward, no GPU, no new seed): the per-prompt rows (n_emitted, n_blocks, n_drafted) of the four committed vanilla gamma-4 arm-1 summaries — Llama-3.1-8B-Instruct NF4 target + Llama-3.2-1B-Instruct bf16 draft, temperature 1.0, 256 new tokens, the HF adapter (acceptance-valid only, D-9); … sideband/FINDINGS.md:302 · d147e8c0cd6b
The full regime line

REAL MODELS (a CPU recompute on committed evidence, no model forward, no GPU, no new seed): the per-prompt rows (n_emitted, n_blocks, n_drafted) of the four committed vanilla gamma-4 arm-1 summaries — Llama-3.1-8B-Instruct NF4 target + Llama-3.2-1B-Instruct bf16 draft, temperature 1.0, 256 new tokens, the HF adapter (acceptance-valid only, D-9); POPULATION K3 = data/prompts_k3_verdict.jsonl at seeds 0 and 1 (results/real/k3_verdict_arm1.summary.json, k3_verdict_s1_arm1.summary.json; F-19's seed pair, THE label population), POPULATION OQ15 = data/prompts_oq15_verdict.jsonl at seeds 2 and 3 (oq15c_verdict_arm1.summary.json, oq15c_verdict_s3_arm1.summary.json; the replication on different prompts and seeds); families instruct / story / summarize, 70 each, bound by the metas' prompt-file shas. Registered 2026-10-04 (docs/BLUEPRINT.md, "SEED-NOISE FLOOR READOUTS REGISTRATION", written before any readout of a seed pair ran, three critic rounds before the freeze), frozen at 906edfd; both runs from the clean tree at that commit. Estimand E6's: tokens/call = sum n_emitted / sum n_blocks on a prompt set; the null delta = the lower seed minus the higher on the same prompts; CI95 by sideband.stats.paired_ratio_delta_ci (n_boot 2000, seed 0), exactly as F-19 and F-21 compute a gate leg; 1,000 subsets without replacement per n below 210, the full set at 210. NON-CRITERION for every finding and kill status: a CROSS-SEED, same-policy, SINGLE-COMPARATOR null, while the record's gate legs are same-seed (D-6) and all-comparator; that same-seed null is not measured (the cross-seed noise is plausibly an upper bound on it, argued from the shared prefix for a policy whose per-prompt variance does not exceed vanilla's; not measured).

Source
results/real/noise_floor_k3.json, results/real/noise_floor_k3.json.state.jsonl (K3, the label population), results/real/noise_floor_oq15.json, results/real/noise_floor_oq15.json.state.jsonl (OQ15, the replication); each state file holds one row per (n, subset) with the subset's prompt indices and swap mask, and the header binds the four summary and meta shas and the prompt file's sha; …
The full source line

results/real/noise_floor_k3.json, results/real/noise_floor_k3.json.state.jsonl (K3, the label population), results/real/noise_floor_oq15.json, results/real/noise_floor_oq15.json.state.jsonl (OQ15, the replication); each state file holds one row per (n, subset) with the subset's prompt indices and swap mask, and the header binds the four summary and meta shas and the prompt file's sha; inputs results/real/k3_verdict_arm1.summary.json, results/real/k3_verdict_s1_arm1.summary.json, results/real/oq15c_verdict_arm1.summary.json, results/real/oq15c_verdict_s3_arm1.summary.json with their .meta.json, data/prompts_k3_verdict.jsonl, data/prompts_oq15_verdict.jsonl; the F-19 val numbers from results/real/k3_threshold_search_arm4.json and results/real/k3_val_arm1.summary.json; instrument tools/run_noise_floor.py (attested by both payloads; frozen); landing test tests/test_noise_floor_landing.py; the registration block and its OUTCOME in docs/BLUEPRINT.md; the methods note docs/NOISE_FLOOR.md.

sideband/FINDINGS.md:303 · d147e8c0cd6b
Date
2026-10-04 sideband/FINDINGS.md:301 · d147e8c0cd6b
Index headline
f_one at n = 30 0.1060 (SE 0.0065) / 0.1320 (SE 0.0070); r 0.4232 [0.3073, 0.5193] / 0.4172 [0.2990, 0.5224]; half-width 0.0385 / 0.0398 at n = 210; g_one at n = 30 0.0235 / 0.0375
0.0385pooled tokens/call by a CI95 half-width of 0.0385 / 0.0398 at 210 promptssideband/FINDINGS.md:300 · d147e8c0cd6b
0.0398by a CI95 half-width of 0.0385 / 0.0398 at 210 prompts andsideband/FINDINGS.md:300 · d147e8c0cd6b
0.0982of 0.0385 / 0.0398 at 210 prompts and 0.0982 / 0.1032 at 30 (K3 /sideband/FINDINGS.md:300 · d147e8c0cd6b
0.10320.0385 / 0.0398 at 210 prompts and 0.0982 / 0.1032 at 30 (K3 / OQ15); onesideband/FINDINGS.md:300 · d147e8c0cd6b
0.4232seed's per-prompt tokens/call correlates 0.4232 / 0.4172 with the othersideband/FINDINGS.md:300 · d147e8c0cd6b
0.4172per-prompt tokens/call correlates 0.4232 / 0.4172 with the other seed's;sideband/FINDINGS.md:300 · d147e8c0cd6b
+0.062/ 0.4172 with the other seed's; at n = 30 a +0.062 point win appears bysideband/FINDINGS.md:300 · d147e8c0cd6b
0.1060point win appears by seed noise alone in 0.1060 / 0.1320 of subsets andsideband/FINDINGS.md:300 · d147e8c0cd6b
sideband/FINDINGS.md:300 · d147e8c0cd6b sideband/docs/FINDINGS_INDEX.md:128 · aa88354d13cb
F-28READER-AND-LENS RE-TEST AT S3, STAGE LENS (the lane's entrysidebandindex: standsABSORBEDOFF3-ABSORBEDOFF4-ABSORBEDSPLITSURVIVES

the full-free re-test reads ABSORBED through the balanced lens where the lopsided lens read SPLIT; F-17's proxy margin survives the lens as existence (LENS-SURVIVES), its growth over L2 NOT-SETTLED

Heading, as written READER-AND-LENS RE-TEST AT S3, STAGE LENS (the lane's entry; later stages add dated lines): through a size-capped BALANCED lens at S3 L3 the full-free re-test reads ABSORBED at the registered reader — gate D +0.0019 [+0.0006, +0.0032] / +0.0004 [-0.0011, +0.0018], every D CI95 upper below +0.005 at off2 / off3 / off4 on both harvests (OFF3-ABSORBED, OFF4-ABSORBED) — where the lopsided lens read SPLIT (F-25); F-17's (c) - (a) proxy margin SURVIVES the lens at +0.0221 [+0.0187, +0.0258] / +0.0195 [+0.0160, +0.0232] (0.307 / 0.267 of the recorded +0.0721 / +0.0730) with its growth over L2 NOT-SETTLED on both harvests (LENS-SURVIVES; combined L3-GROWTH-NOT-SETTLED)

MECHANICAL LABELS (the frozen OQ-18 rules per (reader, space), the primary harvest the label harvest, the replication computed identically; both needed). sideband/FINDINGS.md:332 · d147e8c0cd6b

In plain words
ABSORBED
no material margin over the full free baseline at the registered reader, and no more than a scrambled block of the same size adds; the gate fired on neither harvest of the source
OFF3-ABSORBED
the same absorbed reading three positions ahead instead of two
OFF4-ABSORBED
the same absorbed reading four positions ahead
SPLIT
the gate fired on one of the source's two harvests and not on the other
SURVIVES
the margin cleared its floor under the registered rule
NON-CRITERION
a diagnostic that reports and prices but cannot change a verdict
index: stands
the registered verdict is unchanged and no later entry narrows it
Regime
F-25's — the two committed OLMoE S3 harvests (NF4 double-quant with bf16 routers at harvest time, target-only generations at temperature 1.0, 256 new tokens, 210 prompts, split prompt_idx % 7, 30 test clusters; the sorted top-8 id stream as the 1024-dim slots block) at the REGISTERED reader R-0 (mlp256, 30 epochs, lr 3e-3) and, for the proxy, linear; … sideband/FINDINGS.md:311 · d147e8c0cd6b
The full regime line

F-25's — the two committed OLMoE S3 harvests (NF4 double-quant with bf16 routers at harvest time, target-only generations at temperature 1.0, 256 new tokens, 210 prompts, split prompt_idx % 7, 30 test clusters; the sorted top-8 id stream as the 1024-dim slots block) at the REGISTERED reader R-0 (mlp256, 30 epochs, lr 3e-3) and, for the proxy, linear; probe training only on CPU (six torch threads, CUDA hidden), no model forward, no acceptance, no wall clock; the frozen OQ-18 chain imported by bare name; every label read through the BALANCED L3 lens L3_olmoe_balanced_v1 (16 groups of 31 or 32 concepts; cap 32 against the lopsided lens's 268).

Source
results/real/lens_s3.json, results/real/lens_s3.json.state.jsonl (PRIMARY, harvest_moe), results/real/lens_s3_r.json, results/real/lens_s3_r.json.state.jsonl (REPLICATION, harvest_moe_r), each with 49 rows (45 unshuffled cells, 4 amended-rider cells with histograms); results/real/concept_groups_olmoe_balanced_v1.json (the lens); …
The full source line

results/real/lens_s3.json, results/real/lens_s3.json.state.jsonl (PRIMARY, harvest_moe), results/real/lens_s3_r.json, results/real/lens_s3_r.json.state.jsonl (REPLICATION, harvest_moe_r), each with 49 rows (45 unshuffled cells, 4 amended-rider cells with histograms); results/real/concept_groups_olmoe_balanced_v1.json (the lens); tools/run_reader_lens.py (the instrument, d1125670be05 as frozen), tools/build_balanced_groups.py, tests/test_reader_lens.py, tests/test_balanced_groups.py, tests/test_reader_lens_landing.py; docs/BLUEPRINT.md ("READER-AND-LENS RE-TEST AT S3 REGISTRATION", frozen at b128fda, both runs from the clean tree at d3504b2, the OUTCOME bullet for this stage).

sideband/FINDINGS.md:321 · d147e8c0cd6b
Date
2026-10-05 sideband/FINDINGS.md:310 · d147e8c0cd6b
Index headline
gate D +0.0019 [+0.0006, +0.0032] / +0.0004 [-0.0011, +0.0018]; Mt +0.0042 [+0.0025, +0.0061] / +0.0021 [+0.0002, +0.0041]; Mn_r +0.0023 [+0.0010, +0.0038] / +0.0017 [-0.0000, +0.0035]; P +0.0221 [+0.0187, +0.0258] / +0.0195 [+0.0160, +0.0232] (0.307 / 0.267 of +0.0721 / +0.0730); …
The full headline

gate D +0.0019 [+0.0006, +0.0032] / +0.0004 [-0.0011, +0.0018]; Mt +0.0042 [+0.0025, +0.0061] / +0.0021 [+0.0002, +0.0041]; Mn_r +0.0023 [+0.0010, +0.0038] / +0.0017 [-0.0000, +0.0035]; P +0.0221 [+0.0187, +0.0258] / +0.0195 [+0.0160, +0.0232] (0.307 / 0.267 of +0.0721 / +0.0730); L2 +0.0206 / +0.0207

+0.0019 [+0.0006, +0.0032]ABSORBED at the registered reader — gate D +0.0019 [+0.0006, +0.0032] / +0.0004 [-0.0011,sideband/FINDINGS.md:309 · d147e8c0cd6b
+0.0004 [-0.0011, +0.0018]— gate D +0.0019 [+0.0006, +0.0032] / +0.0004 [-0.0011, +0.0018], every D CI95 uppersideband/FINDINGS.md:309 · d147e8c0cd6b
+0.005+0.0018], every D CI95 upper below +0.005 at off2 / off3 / off4sideband/FINDINGS.md:309 · d147e8c0cd6b
+0.0221 [+0.0187, +0.0258](c) - (a) proxy margin SURVIVES the lens at +0.0221 [+0.0187, +0.0258] / +0.0195 [+0.0160,sideband/FINDINGS.md:309 · d147e8c0cd6b
+0.0195 [+0.0160, +0.0232]the lens at +0.0221 [+0.0187, +0.0258] / +0.0195 [+0.0160, +0.0232] (0.307 / 0.267 of thesideband/FINDINGS.md:309 · d147e8c0cd6b
0.307+0.0258] / +0.0195 [+0.0160, +0.0232] (0.307 / 0.267 of the recordedsideband/FINDINGS.md:309 · d147e8c0cd6b
0.267/ +0.0195 [+0.0160, +0.0232] (0.307 / 0.267 of the recorded +0.0721sideband/FINDINGS.md:309 · d147e8c0cd6b
+0.0721+0.0232] (0.307 / 0.267 of the recorded +0.0721 / +0.0730) with itssideband/FINDINGS.md:309 · d147e8c0cd6b
sideband/FINDINGS.md:309 · d147e8c0cd6b sideband/docs/FINDINGS_INDEX.md:129 · aa88354d13cb
F-29S1 READER RE-TEST: the S1 kill is NOT OVERTURNED under the three F-28 readers — at the S1 gate cell where K-2 …sidebandindex: standsNOT OVERTURNEDABSORBEDK-2-NOT-OVERTURNED-AT-S1NOT-SHOWNK-2-ISOLATED

R-N (hz-rescaled) ABSORBED; R-T (two-tower) ABSORBED; R-L (longer budget) ABSORBED; the S1 word K-2-NOT-OVERTURNED-AT-S1

Heading, as written S1 READER RE-TEST: the S1 kill is NOT OVERTURNED under the three F-28 readers — at the S1 gate cell where K-2 fired at the registered reader (F-24), the hz-rescaled, two-tower and longer-budget readers read R-N (hz-rescaled) ABSORBED (D +0.0026 [+0.0008, +0.0043] / +0.0022 [+0.000497, +0.0039]); R-T (two-tower) ABSORBED (D +0.000339 [-0.000377, +0.0011] / +0.0011 [+0.000363, +0.0019]); R-L (longer budget) ABSORBED (D +0.000382 [-0.0009, +0.0016] / -0.000044 [-0.0014, +0.0013]); the S1 word K-2-NOT-OVERTURNED-AT-S1 (under readers R-N, R-T and R-L at S1; one more family of readers, not K-2's universal)

MECHANICAL LABELS (the frozen OQ-18 rules per reader at the S1 gate cell, both harvests, the primary the label harvest): R-N (hz-rescaled) ABSORBED; R-T (two-tower) ABSORBED; R-L (longer budget) ABSORBED. sideband/FINDINGS.md:870 · d147e8c0cd6b

In plain words
ABSORBED
no material margin over the full free baseline at the registered reader, and no more than a scrambled block of the same size adds; the gate fired on neither harvest of the source
K-2-NOT-OVERTURNED-AT-S1
the S1 kill stands under the stronger readers
NOT-SHOWN
the test could not show the effect; a non-result, not a negative result
K-2-ISOLATED
a stronger reader found a margin at two cells of the lopsided lens, which does not reopen the kill
index: stands
the registered verdict is unchanged and no later entry narrows it
Regime
REAL MODELS (probe training only on the two committed S1 Phase-3 harvests — primary seeds 0..7 results/real/harvest_shadow and replication seeds 8..11 harvest_shadow_r of the NF4 Llama-3.1-8B-Instruct shadow-tap harvest; split prompt_idx % 7, 30 test clusters; no model forward, no acceptance, no wall clock) at the three F-28 readers R-N (hz-rescaled) (mlp256; … sideband/FINDINGS.md:835 · d147e8c0cd6b
The full regime line

REAL MODELS (probe training only on the two committed S1 Phase-3 harvests — primary seeds 0..7 results/real/harvest_shadow and replication seeds 8..11 harvest_shadow_r of the NF4 Llama-3.1-8B-Instruct shadow-tap harvest; split prompt_idx % 7, 30 test clusters; no model forward, no acceptance, no wall clock) at the three F-28 readers R-N (hz-rescaled) (mlp256; 30 epochs, lr 3e-3), R-T (two-tower) (capacity tag mlp256; towers of 256 and 64 units, head over their concatenation; 30 epochs, lr 3e-3), R-L (longer budget) (mlp256; 60 epochs, lr 1e-3), each read at its best VAL epoch, on CPU at six torch threads with CUDA hidden; the frozen reader-lens chain (tools/run_reader_lens.py d1125670be05, attested by six earlier payloads) imported by bare name and never edited; the arms F-24's at the S1 gate cell L1_concept_k16 / offset 2 / mlp256 (fplus = (a) 657 summaries ++ TOK ++ z-scored h_a, 6801 dims; fplus_b = fplus ++ the aligned 2048-dim slots block; fplus_r = fplus ++ the prompt-level-deranged corrected twin, 8849 dims each); readouts the Phase-3 pairing (per-prompt successes summed over three train seeds, paired_cluster_delta_ci over the 30 test prompts, n_boot 2000, seed 0); the +0.005 materiality floor; the frozen OQ-18 rules per reader. Registered 2026-10-16 in docs/BLUEPRINT.md "S1 READER RE-TEST REGISTRATION" before any reader-variant cell at S1 trained, critic-gated through three design rounds and a code review, frozen at 46a812f (the launch commit); CRITERION-BEARING for K-2's status at S1 and for F-24's S1 cells through the pinned consequence rules; NON-CRITERION for everything else. A NEW registration beside the READER-AND-LENS block, whose conditional STAGE S1 stays unrun and whose S3 word does not move.

Source
results/real/readers_s1.json, readers_s1.json.state.jsonl, readers_s1_r.json, readers_s1_r.json.state.jsonl (the two registered runs, 36 rows each), results/real/readers_s1_word.json (the S1 word, written by tools/run_reader_s1.py --conclude at the launch HEAD), tools/run_reader_lens.py (d1125670be05 as frozen), …
The full source line

results/real/readers_s1.json, readers_s1.json.state.jsonl, readers_s1_r.json, readers_s1_r.json.state.jsonl (the two registered runs, 36 rows each), results/real/readers_s1_word.json (the S1 word, written by tools/run_reader_s1.py --conclude at the launch HEAD), tools/run_reader_lens.py (d1125670be05 as frozen), tools/run_reader_s1.py (94640a09cbeb, frozen from this payload), tests/test_reader_s1.py, tests/test_reader_s1_landing.py, docs/BLUEPRINT.md (the S1 READER RE-TEST REGISTRATION and its OUTCOME); the R-0 reference rows results/real/fullfree_s1.json, fullfree_s1_r.json, fullfree_s1_twin.json, fullfree_s1_r_twin.json and their .state.jsonl (F-24).

sideband/FINDINGS.md:860 · d147e8c0cd6b
Date
2026-10-17 sideband/FINDINGS.md:834 · d147e8c0cd6b
Index headline
R-N (hz-rescaled) D +0.0026 [+0.0008, +0.0043] / +0.0022 [+0.000497, +0.0039], Mt -0.0012 [-0.0028, +0.000459] / -0.0028 [-0.0058, -0.000142], Mn_r -0.0037 [-0.0054, -0.0021] / -0.0050 [-0.0073, -0.0028]; …
The full headline

R-N (hz-rescaled) D +0.0026 [+0.0008, +0.0043] / +0.0022 [+0.000497, +0.0039], Mt -0.0012 [-0.0028, +0.000459] / -0.0028 [-0.0058, -0.000142], Mn_r -0.0037 [-0.0054, -0.0021] / -0.0050 [-0.0073, -0.0028]; R-T (two-tower) D +0.000339 [-0.000377, +0.0011] / +0.0011 [+0.000363, +0.0019], Mt +0.000000 [-0.0013, +0.0014] / +0.0059 [+0.0039, +0.0079], Mn_r -0.000339 [-0.0017, +0.0010] / +0.004847 [+0.0030, +0.0067]; R-L (longer budget) D +0.000382 [-0.0009, +0.0016] / -0.000044 [-0.0014, +0.0013], Mt +0.0031 [+0.0019, +0.0043] / +0.0023 [+0.0007, +0.0039], Mn_r +0.0027 [+0.0011, +0.0042] / +0.0023 [+0.000415, +0.0043]

+0.0026 [+0.0008, +0.0043]readers read R-N (hz-rescaled) ABSORBED (D +0.0026 [+0.0008, +0.0043] / +0.0022 [+0.000497,sideband/FINDINGS.md:833 · d147e8c0cd6b
+0.0022 [+0.000497, +0.0039]ABSORBED (D +0.0026 [+0.0008, +0.0043] / +0.0022 [+0.000497, +0.0039]); R-T (two-tower)sideband/FINDINGS.md:833 · d147e8c0cd6b
+0.000339 [-0.000377, +0.0011]+0.0039]); R-T (two-tower) ABSORBED (D +0.000339 [-0.000377, +0.0011] / +0.0011 [+0.000363,sideband/FINDINGS.md:833 · d147e8c0cd6b
+0.0011 [+0.000363, +0.0019](D +0.000339 [-0.000377, +0.0011] / +0.0011 [+0.000363, +0.0019]); R-L (longer budget)sideband/FINDINGS.md:833 · d147e8c0cd6b
+0.000382 [-0.0009, +0.0016]+0.0019]); R-L (longer budget) ABSORBED (D +0.000382 [-0.0009, +0.0016] / -0.000044 [-0.0014,sideband/FINDINGS.md:833 · d147e8c0cd6b
-0.000044 [-0.0014, +0.0013]ABSORBED (D +0.000382 [-0.0009, +0.0016] / -0.000044 [-0.0014, +0.0013]); the S1 wordsideband/FINDINGS.md:833 · d147e8c0cd6b
+0.0026 [+0.0008, +0.0043]udget readers: R-N (hz-rescaled) ABSORBED (D +0.0026 [+0.0008,sideband/FINDINGS.md:877 · d147e8c0cd6b
+0.0022 [+0.000497, +0.0039]+0.0043] / +0.0022 [+0.000497, +0.0039]); R-T (two-tower)sideband/FINDINGS.md:878 · d147e8c0cd6b
sideband/FINDINGS.md:833 · d147e8c0cd6b sideband/docs/FINDINGS_INDEX.md:130 · aa88354d13cb
F-1v0 codec is byte-exact; beats the draft-AC bound where alignment is decent; loses to content-level coding in …amber

Heading, as written v0 codec is byte-exact; beats the draft-AC bound where alignment is decent; loses to content-level coding in this degenerate world

decode(encode(x)) == x exactly across configs and seeds (17 tests). amber/FINDINGS.md:10 · 711fe31da2a4

Regime
sideband synthetic concept-HMM v0 world, TARGET-GREEDY content, amber v0 rank codec (Elias-gamma gaps+ranks), T=20k x 2 seeds; headers excluded from rates amber/FINDINGS.md:8 · 711fe31da2a4
Source
  • results/a01_rate_vs_draft_quality.json
  • experiments/a01_rate_vs_draft_quality.py
  • tests/test_amber.py
amber/FINDINGS.md:9 · 711fe31da2a4
Date
2026-08-28 amber/FINDINGS.md:7 · 711fe31da2a4
0.5and seeds (17 tests). Rates by draft_alpha: 0.5 -> 9.29 bits/tokenamber/FINDINGS.md:10 · 711fe31da2a4
9.29(17 tests). Rates by draft_alpha: 0.5 -> 9.29 bits/token (WORSE thanamber/FINDINGS.md:10 · 711fe31da2a4
9.130.5 -> 9.29 bits/token (WORSE than raw 9.13 — the format losesamber/FINDINGS.md:10 · 711fe31da2a4
0.55than raw 9.13 — the format loses below ~0.55 alignment); 0.7 -> 5.84amber/FINDINGS.md:10 · 711fe31da2a4
0.7— the format loses below ~0.55 alignment); 0.7 -> 5.84 (beats theamber/FINDINGS.md:10 · 711fe31da2a4
5.84format loses below ~0.55 alignment); 0.7 -> 5.84 (beats theamber/FINDINGS.md:10 · 711fe31da2a4
6.490.7 -> 5.84 (beats the AC-under-draft bound 6.49 — regime-A structureamber/FINDINGS.md:10 · 711fe31da2a4
0.9structure the draft alone cannot see); 0.9 -> 2.17 (2.3x the ACamber/FINDINGS.md:10 · 711fe31da2a4
amber/FINDINGS.md:6 · 711fe31da2a4
F-2Attestation works at L1; the L0 blind spot is exactly order-preserving drift — which is also a featureamber

honest archive attests exactly (recomputed divergences == recorded corrections, 2346/2346); a single forged token fails L1. amber/FINDINGS.md:17 · 711fe31da2a4

Regime
amber v0 rank codec on the sideband synthetic world, T=8000; L0 = header fingerprint, L1 = target divergence-set equality; substitutions bypass L0 deliberately amber/FINDINGS.md:15 · 711fe31da2a4
Source
  • results/a02_attestation.json
  • experiments/a02_attestation.py
  • tests/test_amber.py
amber/FINDINGS.md:16 · 711fe31da2a4
Date
2026-08-28 amber/FINDINGS.md:14 · 711fe31da2a4
2346/2346divergences == recorded corrections, 2346/2346); a single forged tokenamber/FINDINGS.md:17 · 711fe31da2a4
0.7substitutions with L0 bypassed: draft_alpha 0.7->0.6 changes 49.3% ofamber/FINDINGS.md:17 · 711fe31da2a4
0.6with L0 bypassed: draft_alpha 0.7->0.6 changes 49.3% ofamber/FINDINGS.md:17 · 711fe31da2a4
49.3%L0 bypassed: draft_alpha 0.7->0.6 changes 49.3% of decoded tokensamber/FINDINGS.md:17 · 711fe31da2a4
1.649.3% of decoded tokens (loud); temperature 1.6->2.4 changes 0.0% andamber/FINDINGS.md:17 · 711fe31da2a4
2.4of decoded tokens (loud); temperature 1.6->2.4 changes 0.0% and zipf_samber/FINDINGS.md:17 · 711fe31da2a4
0.0%tokens (loud); temperature 1.6->2.4 changes 0.0% and zipf_s 1.1->1.2amber/FINDINGS.md:17 · 711fe31da2a4
1.11.6->2.4 changes 0.0% and zipf_s 1.1->1.2 changes 0.0% —amber/FINDINGS.md:17 · 711fe31da2a4
amber/FINDINGS.md:13 · 711fe31da2a4
F-3Rank-census instrument validated on synthetic driftamber

Heading, as written Rank-census instrument validated on synthetic drift: flips obey the Gaussian near-tie law, are fully attributable to near-ties, and this world has a constant top-2 gap

top-1 flip rate vs the pre-registered law mean_t 0.5*erfc(gap_t/2eps): eps=0.05 -> 0.0041 measured [0.0030, 0.0052 across seeds] vs 0.0042 predicted; eps=0.1 -> 0.0957 [0.0932, 0.0973] vs 0.0940; eps=0.2 -> 0.278 vs 0.255 and eps=0.3 -> 0.396 vs 0.330 (law is a stated lower bound: it ignores rank-3+ intruders, which appear once eps ~ gap). amber/FINDINGS.md:25 · 711fe31da2a4

Regime
sideband synthetic concept-HMM v0 world drafts (alpha 0.7, T=1.6, V=560), i.i.d. Gaussian LOGIT-space drift eps in {1e-3..3e-1}, T=6000 positions x 3 noise seeds; tau calibrated at the 99.9th pct of a benign draw at the same eps; instrument shakedown only, not a hardware measurement amber/FINDINGS.md:23 · 711fe31da2a4
Source
  • results/a03_rank_flip_synthetic.json
  • experiments/a03_rank_flip_synthetic.py
  • src/amber/rankcensus.py
  • tests/test_rankcensus.py
amber/FINDINGS.md:24 · 711fe31da2a4
Date
2026-08-29 amber/FINDINGS.md:22 · 711fe31da2a4
0.5flip rate vs the pre-registered law mean_t 0.5*erfc(gap_t/2eps):amber/FINDINGS.md:25 · 711fe31da2a4
0.05law mean_t 0.5*erfc(gap_t/2eps): eps=0.05 -> 0.0041 measuredamber/FINDINGS.md:25 · 711fe31da2a4
0.0041mean_t 0.5*erfc(gap_t/2eps): eps=0.05 -> 0.0041 measured [0.0030,amber/FINDINGS.md:25 · 711fe31da2a4
0.0030eps=0.05 -> 0.0041 measured [0.0030, 0.0052 across seeds]amber/FINDINGS.md:25 · 711fe31da2a4
0.0052eps=0.05 -> 0.0041 measured [0.0030, 0.0052 across seeds] vs 0.0042amber/FINDINGS.md:25 · 711fe31da2a4
0.0042measured [0.0030, 0.0052 across seeds] vs 0.0042 predicted; eps=0.1 ->amber/FINDINGS.md:25 · 711fe31da2a4
0.1across seeds] vs 0.0042 predicted; eps=0.1 -> 0.0957 [0.0932,amber/FINDINGS.md:25 · 711fe31da2a4
0.0957 [0.0932, 0.0973]seeds] vs 0.0042 predicted; eps=0.1 -> 0.0957 [0.0932, 0.0973] vs 0.0940; eps=0.2 ->amber/FINDINGS.md:25 · 711fe31da2a4
amber/FINDINGS.md:21 · 711fe31da2a4
F-4The seeded mutation sweep found silent header corruptionamber

Heading, as written The seeded mutation sweep found silent header corruption: a one-bit flip in the length digits decoded 'successfully' to a longer sequence; fixed by sealing the header

v0.1.0's L0 fingerprint covered only (cfg, path_seed, draft_seed); length, n_corrections and payload_bits were unauthenticated, so an ASCII bit flip turning "300" into "310" parsed, validated, replayed and returned 310 tokens with no error — the exact silent-success class lockstep F-3 found in its container. amber/FINDINGS.md:32 · 711fe31da2a4

Regime
amber v0.1.0 archive framing (JSON header + Elias-gamma payload), 400 DRBG-seeded single-byte mutations + exhaustive 1-bit flips over every header byte, T=300 synthetic archive amber/FINDINGS.md:30 · 711fe31da2a4
Source
tests/test_hardening.py, src/amber/codec.py (validate_header, header_digest) amber/FINDINGS.md:31 · 711fe31da2a4
Date
2026-08-29 amber/FINDINGS.md:29 · 711fe31da2a4
300so an ASCII bit flip turning "300" into "310" parsed,amber/FINDINGS.md:32 · 711fe31da2a4
310so an ASCII bit flip turning "300" into "310" parsed, validated,amber/FINDINGS.md:32 · 711fe31da2a4
310parsed, validated, replayed and returned 310 tokens with no error —amber/FINDINGS.md:32 · 711fe31da2a4
16(exact field set, int-not-bool, ranges, 16-hex digests),amber/FINDINGS.md:32 · 711fe31da2a4
2fields, declared-length ceiling (default 2^24, caller-tunable)amber/FINDINGS.md:32 · 711fe31da2a4
400duplicate-key and non-finite rejection; 400-mutation sweep +amber/FINDINGS.md:32 · 711fe31da2a4
amber/FINDINGS.md:28 · 711fe31da2a4
F-5Phase 2 rank census (TinyLlama-1.1B, one 3080 Ti box)amber

Heading, as written Phase 2 rank census (TinyLlama-1.1B, one 3080 Ti box): fp32 cross-device rank order is exactly stable; bf16 flips 1.0% of top-1s, every one a near-tie; run-to-run is bit-identical; the pre-registered band (c) fired on bf16 batch dependence, which the investigation attributes to kernel selection (N2), not a bug

vs CPU fp32 — cuda_fp32: 0 top-1 flips / 49,152 (top-64 SET flip 1.4e-4, p99 |dlogit| 1.5e-5, max 1.3e-4); cuda_fp16: 0.114% (56 flips; gap median 0.0044, max 0.041; 0 beyond 2tau=0.142); cuda_bf16: 1.015% (499; gap median 0.028, max 0.23; 0 beyond 2tau=0.93); int8: 4.04% (0 beyond 2tau=5.0); nf4: 10.2% (3 beyond 2tau=12.3, gap max 7.06 — quantization changed the MODEL, not drift). amber/FINDINGS.md:40 · 711fe31da2a4

Regime
TinyLlama/TinyLlama-1.1B-Chat-v1.0 (Llama arch, V=32000; snapshot fe8a4ea1) teacher-forced on 96x512 windows of tokenizer-lab mix_docs50_wiki50 (49,152 positions, no padding); top-256 dumps, census at k=64; reference CPU fp32; RTX 3080 Ti, torch 2.6.0+cu124, transformers 4.57.3, cudnn default (deterministic_algorithms off); envprobe 2d2a2e0c; bands pre-registered in experiments/P2_PROTOCOL.md. amber/FINDINGS.md:38 · 711fe31da2a4
The full regime line

TinyLlama/TinyLlama-1.1B-Chat-v1.0 (Llama arch, V=32000; snapshot fe8a4ea1) teacher-forced on 96x512 windows of tokenizer-lab mix_docs50_wiki50 (49,152 positions, no padding); top-256 dumps, census at k=64; reference CPU fp32; RTX 3080 Ti, torch 2.6.0+cu124, transformers 4.57.3, cudnn default (deterministic_algorithms off); envprobe 2d2a2e0c; bands pre-registered in experiments/P2_PROTOCOL.md. NOT Llama-3.2-1B (gated repo not authorized on this box); same command re-runs it when access lands.

Source
  • results/p2_rank_census.json
  • experiments/p2_rank_census.py
  • experiments/P2_PROTOCOL.md
  • src/amber/real/capture.py
  • src/amber/rankcensus.py
amber/FINDINGS.md:39 · 711fe31da2a4
Date
2026-08-29 amber/FINDINGS.md:37 · 711fe31da2a4
1.0%rank order is exactly stable; bf16 flips 1.0% of top-1s, every one aamber/FINDINGS.md:36 · 711fe31da2a4
1.4e-40 top-1 flips / 49,152 (top-64 SET flip 1.4e-4, p99 |dlogit| 1.5e-5,amber/FINDINGS.md:40 · 711fe31da2a4
1.5e-5(top-64 SET flip 1.4e-4, p99 |dlogit| 1.5e-5, max 1.3e-4);amber/FINDINGS.md:40 · 711fe31da2a4
1.3e-4SET flip 1.4e-4, p99 |dlogit| 1.5e-5, max 1.3e-4); cuda_fp16: 0.114% (56amber/FINDINGS.md:40 · 711fe31da2a4
0.114%|dlogit| 1.5e-5, max 1.3e-4); cuda_fp16: 0.114% (56 flips; gap medianamber/FINDINGS.md:40 · 711fe31da2a4
0.0044cuda_fp16: 0.114% (56 flips; gap median 0.0044, max 0.041; 0 beyondamber/FINDINGS.md:40 · 711fe31da2a4
0.0410.114% (56 flips; gap median 0.0044, max 0.041; 0 beyond 2tau=0.142);amber/FINDINGS.md:40 · 711fe31da2a4
0.142gap median 0.0044, max 0.041; 0 beyond 2tau=0.142); cuda_bf16: 1.015%amber/FINDINGS.md:40 · 711fe31da2a4
amber/FINDINGS.md:36 · 711fe31da2a4
F-6Same dumps read as lockstep's E-006amber

Heading, as written Same dumps read as lockstep's E-006: fp32 cross-device drift crosses a 1e-3 logit grid at 22% of positions (law k*eps/delta holds), bf16/fp16 at 100% — rank stability is orders of magnitude cheaper than grid stability

CPU fp32 vs GPU fp32: eps p50 3e-6, p99 1.5e-5, max 1.3e-4 (absolute, on logits of magnitude ~10); positions with >=1 crossing: 2.6% at delta=1e-2, 21.9% at delta=1e-3, 90.8% at delta=1e-4. amber/FINDINGS.md:48 · 711fe31da2a4

Regime
as F-5; crossing = any CPU-top-64 token changes floor(logit/delta) or leaves the GPU arm's top-64; delta in {1e-2, 1e-3, 1e-4}, k=64 (lockstep D-13's proposed profile) amber/FINDINGS.md:46 · 711fe31da2a4
Source
  • results/p2_e006_grid_crossing.json
  • results/p2_rank_census.json
amber/FINDINGS.md:47 · 711fe31da2a4
Date
2026-08-29 amber/FINDINGS.md:45 · 711fe31da2a4
1e-3E-006: fp32 cross-device drift crosses a 1e-3 logit grid at 22% ofamber/FINDINGS.md:44 · 711fe31da2a4
22%drift crosses a 1e-3 logit grid at 22% of positions (lawamber/FINDINGS.md:44 · 711fe31da2a4
100%(law k*eps/delta holds), bf16/fp16 at 100% — rank stability isamber/FINDINGS.md:44 · 711fe31da2a4
3e-6- finding: CPU fp32 vs GPU fp32: eps p50 3e-6, p99 1.5e-5, max 1.3e-4amber/FINDINGS.md:48 · 711fe31da2a4
1.5e-5CPU fp32 vs GPU fp32: eps p50 3e-6, p99 1.5e-5, max 1.3e-4 (absolute,amber/FINDINGS.md:48 · 711fe31da2a4
1.3e-4vs GPU fp32: eps p50 3e-6, p99 1.5e-5, max 1.3e-4 (absolute, on logits ofamber/FINDINGS.md:48 · 711fe31da2a4
2.6%~10); positions with >=1 crossing: 2.6% at delta=1e-2, 21.9% atamber/FINDINGS.md:48 · 711fe31da2a4
1e-2positions with >=1 crossing: 2.6% at delta=1e-2, 21.9% at delta=1e-3,amber/FINDINGS.md:48 · 711fe31da2a4
amber/FINDINGS.md:44 · 711fe31da2a4
F-7Phase 2 replicated on the intended draft (Llama-3.2-1B-Instruct)amber

Heading, as written Phase 2 replicated on the intended draft (Llama-3.2-1B-Instruct): same shape, larger magnitudes — fp32 order stable to 1 flip in 49k, fp16 0.42%, bf16 2.0% (all near-ties), run-to-run bit-identical; a 1e-3 grid crosses at 41% of positions even at fp32

vs CPU fp32 — cuda_fp32 b1: 1 top-1 flip / 49,152 (2.0e-5) at a reference gap of 1.3e-5 (an exact near-tie; p99 |dlogit| 2.9e-5, max 4.5e-4); cuda_fp32 b8: 0; fp16: 0.42% (208; gap median 0.0017, max 0.024; 0 beyond 2tau=0.066); bf16: 2.00% (984; gap median 0.032, max 0.263; 0 beyond 2tau=0.93); int8: 5.6% (2 beyond 2tau=6.6); nf4: 16.0%. amber/FINDINGS.md:55 · 711fe31da2a4

Regime
meta-llama/Llama-3.2-1B-Instruct (V=128256; snapshot 9213176726f5) teacher-forced on 96x512 windows of tokenizer-lab mix_docs50_wiki50 re-tokenized with the Llama-3 tokenizer (49,152 positions, no padding; corpus sha differs from F-5's); top-256 dumps, census at k=64; CPU fp32 reference; same box/stack as F-5 (3080 Ti, torch 2.6.0+cu124, transformers 4.57.3); … amber/FINDINGS.md:53 · 711fe31da2a4
The full regime line

meta-llama/Llama-3.2-1B-Instruct (V=128256; snapshot 9213176726f5) teacher-forced on 96x512 windows of tokenizer-lab mix_docs50_wiki50 re-tokenized with the Llama-3 tokenizer (49,152 positions, no padding; corpus sha differs from F-5's); top-256 dumps, census at k=64; CPU fp32 reference; same box/stack as F-5 (3080 Ti, torch 2.6.0+cu124, transformers 4.57.3); P2_PROTOCOL.md unchanged

Source
  • results/p2_llama32_rank_census.json
  • results/p2_llama32_e006_grid_crossing.json
  • experiments/p2_rank_census.py
  • results/p2_llama32.log (capture provenance/throughput)
amber/FINDINGS.md:54 · 711fe31da2a4
Date
2026-08-29 amber/FINDINGS.md:52 · 711fe31da2a4
0.42%— fp32 order stable to 1 flip in 49k, fp16 0.42%, bf16 2.0% (allamber/FINDINGS.md:51 · 711fe31da2a4
2.0%stable to 1 flip in 49k, fp16 0.42%, bf16 2.0% (all near-ties),amber/FINDINGS.md:51 · 711fe31da2a4
1e-3near-ties), run-to-run bit-identical; a 1e-3 grid crosses at 41% ofamber/FINDINGS.md:51 · 711fe31da2a4
41%bit-identical; a 1e-3 grid crosses at 41% of positions even atamber/FINDINGS.md:51 · 711fe31da2a4
2.0e-5fp32 — cuda_fp32 b1: 1 top-1 flip / 49,152 (2.0e-5) at a reference gap ofamber/FINDINGS.md:55 · 711fe31da2a4
1.3e-5/ 49,152 (2.0e-5) at a reference gap of 1.3e-5 (an exact near-tie; p99amber/FINDINGS.md:55 · 711fe31da2a4
2.9e-5of 1.3e-5 (an exact near-tie; p99 |dlogit| 2.9e-5, max 4.5e-4); cuda_fp32amber/FINDINGS.md:55 · 711fe31da2a4
4.5e-4exact near-tie; p99 |dlogit| 2.9e-5, max 4.5e-4); cuda_fp32 b8: 0;amber/FINDINGS.md:55 · 711fe31da2a4
amber/FINDINGS.md:51 · 711fe31da2a4
F-8Phase 1 (OQ-1) on the real Llama-3.2 pairamberPASSAC-ONLY

Heading, as written Phase 1 (OQ-1) on the real Llama-3.2 pair: 6.8% greedy divergence; amber v0 codes target-generated text at 0.47 bits/token — 7.6x below zstd-19 — and ties AC-under-draft exactly; the pre-registered bands had a dead zone there; target generate-vs-forward self-consistency is 99.67%, below the protocol's 99.9% line

divergence 6.78% overall (instruct 5.8%, story 7.0%, summarize 7.5%); per-prompt spread p10 1.2% / p50 5.1% / p90 17.4%; strongly front-loaded — 15.2% in generated positions 0-63, 6.2% at 64-127, 3.5% at 128-191, 1.8% at 192-255; mildly clustered (P(div | prev div) 0.21 vs 0.068 base; runs mean 1.26, max 6). amber/FINDINGS.md:64 · 711fe31da2a4

In plain words
PASS
the registered gate passed
Regime
target meta-llama/Llama-3.2-3B (base, snapshot 13afe512) greedy, bf16, batch 1, KV cache, max_new 256, sideband's first 200 prompts (3 families, short instruction-style); draft meta-llama/Llama-3.2-1B (base, 4e20de36) teacher-forced bf16 full-vocab stable rank; 48,958 generated tokens (193/200 prompts hit the 256 cap, min 21, 0 empty — stop rule not triggered); … amber/FINDINGS.md:62 · 711fe31da2a4
The full regime line

target meta-llama/Llama-3.2-3B (base, snapshot 13afe512) greedy, bf16, batch 1, KV cache, max_new 256, sideband's first 200 prompts (3 families, short instruction-style); draft meta-llama/Llama-3.2-1B (base, 4e20de36) teacher-forced bf16 full-vocab stable rank; 48,958 generated tokens (193/200 prompts hit the 256 cap, min 21, 0 empty — stop rule not triggered); content coders on the decoded UTF-8 text (4.73 bytes/token); headers excluded; one 3080 Ti; P1_PROTOCOL.md frozen before generation

Source
  • results/p1_divergence.json
  • experiments/p1_divergence.py
  • experiments/P1_PROTOCOL.md
  • src/amber/real/divergence.py
  • results/p1.log
amber/FINDINGS.md:63 · 711fe31da2a4
Date
2026-08-29 amber/FINDINGS.md:61 · 711fe31da2a4
6.8%Phase 1 (OQ-1) on the real Llama-3.2 pair: 6.8% greedy divergence;amber/FINDINGS.md:60 · 711fe31da2a4
0.47amber v0 codes target-generated text at 0.47 bits/token — 7.6x belowamber/FINDINGS.md:60 · 711fe31da2a4
99.67%generate-vs-forward self-consistency is 99.67%, below the protocol'samber/FINDINGS.md:60 · 711fe31da2a4
99.9%is 99.67%, below the protocol's 99.9% lineamber/FINDINGS.md:60 · 711fe31da2a4
6.78%- finding: divergence 6.78% overall (instruct 5.8%,amber/FINDINGS.md:64 · 711fe31da2a4
5.8%finding: divergence 6.78% overall (instruct 5.8%, story 7.0%, summarizeamber/FINDINGS.md:64 · 711fe31da2a4
7.0%6.78% overall (instruct 5.8%, story 7.0%, summarize 7.5%);amber/FINDINGS.md:64 · 711fe31da2a4
7.5%(instruct 5.8%, story 7.0%, summarize 7.5%); per-prompt spread p10amber/FINDINGS.md:64 · 711fe31da2a4
amber/FINDINGS.md:60 · 711fe31da2a4
F-9Coder upgrade (lockstep rac, order-only contexts) gains 1% over Elias-gamma on the real stream and PASSES the pre-registered rac-vs-AC band on held-outamberPASSESPASS

Heading, as written Coder upgrade (lockstep rac, order-only contexts) gains 1% over Elias-gamma on the real stream and PASSES the pre-registered rac-vs-AC band on held-out — but the sign of amber-vs-AC flips between prompt halves, the adaptive coder LOSES on 250-token archives, and the only large lever (value contexts, -23%) is the unsafe one; the tie escape costs 0.14-0.42 bits/token

Held-out, concatenated stream (bits/token): gamma v0 0.4213, rac1 0.4171, AC-under-draft 0.4380, zstd-19 3.470, raw 16.97. amber/FINDINGS.md:73 · 711fe31da2a4

In plain words
PASSES
the pre-registered bar was passed
PASS
the registered gate passed
Regime
P1 stream (F-8: Llama-3.2-3B base greedy -> 1B base draft ranks, bf16, 200 sideband prompts x <=256 tokens); tuning rows 0-99 (24,178 tokens, divergence 7.49%), held-out rows 100-199 (24,780 tokens, 6.10%); rates are real payload bytes x 8 per token, headers excluded; coder = lockstep.rac (integer WNC) with amber's small-alphabet adaptive models; hyperparameters (increment 64, rescale 4096 — … amber/FINDINGS.md:71 · 711fe31da2a4
The full regime line

P1 stream (F-8: Llama-3.2-3B base greedy -> 1B base draft ranks, bf16, 200 sideband prompts x <=256 tokens); tuning rows 0-99 (24,178 tokens, divergence 7.49%), held-out rows 100-199 (24,780 tokens, 6.10%); rates are real payload bytes x 8 per token, headers excluded; coder = lockstep.rac (integer WNC) with amber's small-alphabet adaptive models; hyperparameters (increment 64, rescale 4096 — the most adaptive corner of the grid) chosen on the tuning split only; P3_PROTOCOL.md frozen before the coder saw the data

Source
  • results/p3_rac_rate.json
  • experiments/p3_rac_rate.py
  • experiments/P3_PROTOCOL.md
  • src/amber/rac_codec.py
  • tests/test_rac_codec.py
  • results/p3.log
amber/FINDINGS.md:72 · 711fe31da2a4
Date
2026-08-30 amber/FINDINGS.md:70 · 711fe31da2a4
1%(lockstep rac, order-only contexts) gains 1% over Elias-gamma on theamber/FINDINGS.md:69 · 711fe31da2a4
-23%and the only large lever (value contexts, -23%) is the unsafe one; theamber/FINDINGS.md:69 · 711fe31da2a4
0.14-0.42is the unsafe one; the tie escape costs 0.14-0.42 bits/tokenamber/FINDINGS.md:69 · 711fe31da2a4
0.4213concatenated stream (bits/token): gamma v0 0.4213, rac1 0.4171,amber/FINDINGS.md:73 · 711fe31da2a4
0.4171stream (bits/token): gamma v0 0.4213, rac1 0.4171, AC-under-draft 0.4380,amber/FINDINGS.md:73 · 711fe31da2a4
0.4380v0 0.4213, rac1 0.4171, AC-under-draft 0.4380, zstd-19 3.470, rawamber/FINDINGS.md:73 · 711fe31da2a4
3.470rac1 0.4171, AC-under-draft 0.4380, zstd-19 3.470, raw 16.97. Theamber/FINDINGS.md:73 · 711fe31da2a4
16.97AC-under-draft 0.4380, zstd-19 3.470, raw 16.97. The pre-registeredamber/FINDINGS.md:73 · 711fe31da2a4
amber/FINDINGS.md:69 · 711fe31da2a4
F-10The near-tie escape does exactly what it was built for (top-1 desyncs -> 0 at theta >= 2*tau) and it is not enoughamberZERO

Heading, as written The near-tie escape does exactly what it was built for (top-1 desyncs -> 0 at theta >= 2*tau) and it is not enough: on the P2 dumps a bf16 decoder mis-resolves 3% of rank-2 and 9-26% of rank-3..8 corrections written by an fp32 (or batch-1 bf16) encoder, which on P1's rank distribution is one desync per ~170 target-generated tokens; only an fp32-class decoder is clean (0 desyncs / 49k positions / all ranks to 64)

(1) Escape works as specified: A-desyncs fp32->fp16 25 -> 0 at theta 0.142; fp32->bf16 159 -> 2 at 0.142 -> 0 at 0.465; bf16_b1->bf16_b8 115 -> 0 at 0.465; every A-desync at theta=0 has encoder gap below the pair budget 2(tau_enc+tau_dec) (XB-3 check: no wide-gap flip, no pipeline bug). amber/FINDINGS.md:80 · 711fe31da2a4

In plain words
ZERO
no measurable effect; the measured value is zero within its interval
Regime
results/p2_llama32 dumps (Llama-3.2-1B-Instruct, top-256 per arm, 49,056 teacher-forced positions of tokenizer-lab mix_docs50_wiki50 text — real text, NOT target-generated: true-token rank median 1 / p90 80 / p99 257, 10.9% beyond R=64 and coded X); symbol stream built under the encoder arm (cpu_fp32; cuda_bf16_b1 for the serving case) with the v2 classifier, resolved under each decoder arm; … amber/FINDINGS.md:78 · 711fe31da2a4
The full regime line

results/p2_llama32 dumps (Llama-3.2-1B-Instruct, top-256 per arm, 49,056 teacher-forced positions of tokenizer-lab mix_docs50_wiki50 text — real text, NOT target-generated: true-token rank median 1 / p90 80 / p99 257, 10.9% beyond R=64 and coded X); symbol stream built under the encoder arm (cpu_fp32; cuda_bf16_b1 for the serving case) with the v2 classifier, resolved under each decoder arm; counts of positions whose resolved token differs; no coder state involved (the coder's contexts are value-free, so a desync is a wrong token, never a lost stream)

Source
  • results/p2b_escape_robustness.json
  • experiments/p2b_escape_robustness.py
  • experiments/P3_PROTOCOL.md
  • results/p2b.log
  • results/p2_llama32_rank_census.json
amber/FINDINGS.md:79 · 711fe31da2a4
Date
2026-08-30 amber/FINDINGS.md:77 · 711fe31da2a4
3%on the P2 dumps a bf16 decoder mis-resolves 3% of rank-2 and 9-26% ofamber/FINDINGS.md:76 · 711fe31da2a4
9-26%bf16 decoder mis-resolves 3% of rank-2 and 9-26% of rank-3..8amber/FINDINGS.md:76 · 711fe31da2a4
0.142A-desyncs fp32->fp16 25 -> 0 at theta 0.142; fp32->bf16 159 -> 2 atamber/FINDINGS.md:80 · 711fe31da2a4
0.142-> 0 at theta 0.142; fp32->bf16 159 -> 2 at 0.142 -> 0 at 0.465;amber/FINDINGS.md:80 · 711fe31da2a4
0.4650.142; fp32->bf16 159 -> 2 at 0.142 -> 0 at 0.465; bf16_b1->bf16_b8 115amber/FINDINGS.md:80 · 711fe31da2a4
0.465-> 0 at 0.465; bf16_b1->bf16_b8 115 -> 0 at 0.465; every A-desync atamber/FINDINGS.md:80 · 711fe31da2a4
1.3e-5fp32 -> cuda fp32: 1 desync at theta=0 (gap 1.3e-5), ZERO of any class atamber/FINDINGS.md:80 · 711fe31da2a4
0.142(gap 1.3e-5), ZERO of any class at theta >= 0.142, across all ranks to 64amber/FINDINGS.md:80 · 711fe31da2a4
amber/FINDINGS.md:76 · 711fe31da2a4
F-11Real archives decode byte-exactly with the draft alone under the fp32 contract (200/200 on GPU, 24/24 on CPU, incremental KV replay against a teacher-forced encoder)amber

Heading, as written Real archives decode byte-exactly with the draft alone under the fp32 contract (200/200 on GPU, 24/24 on CPU, incremental KV replay against a teacher-forced encoder); bf16 decoders fail 84% of archives and fp16 28.5%, both inside the bands F-10 predicted; on 245-token archives the header, not the payload, is the file size

cuda_fp32: 200/200 archives byte-exact, zero desyncs in 48,958 incrementally replayed positions — the KV-cache incremental path reproduces the teacher-forced full-forward order at every coded depth (the fp32 analogue of F-8's 0.33% bf16 generate-vs-forward mismatch is 0/49k). amber/FINDINGS.md:90 · 711fe31da2a4

Regime
P1 content (Llama-3.2-3B greedy, 200 sideband prompts, 48,958 tokens, prompts mean 13.5 tokens) archived with Llama-3.2-1B base (snapshot 4e20de36) on cuda fp32, ONE teacher-forced full forward per prompt (991 pos/s incl. amber/FINDINGS.md:88 · 711fe31da2a4
The full regime line

P1 content (Llama-3.2-3B greedy, 200 sideband prompts, 48,958 tokens, prompts mean 13.5 tokens) archived with Llama-3.2-1B base (snapshot 4e20de36) on cuda fp32, ONE teacher-forced full forward per prompt (991 pos/s incl. rank/top-k extraction); gamma coder, tie escape off, max_rank 64 (D-8 defaults); decoded draft-only, incremental KV cache, batch 1, canonical top-128 per step, stopping at the first mismatch (truth known); arms cuda_fp32 / cuda_bf16 / cuda_fp16 (200 archives each) and cpu_fp32 (first 24); one 3080 Ti; P4_PROTOCOL.md frozen before any archive was written

Source
  • results/p4_real_archive.json
  • experiments/p4_real_archive.py
  • experiments/P4_PROTOCOL.md
  • src/amber/real/archive.py
  • tests/test_hf_archive.py
  • results/p4_decode.log
  • results/p4_report.log
amber/FINDINGS.md:89 · 711fe31da2a4
Date
2026-08-30 amber/FINDINGS.md:87 · 711fe31da2a4
200/200the draft alone under the fp32 contract (200/200 on GPU, 24/24 on CPU,amber/FINDINGS.md:86 · 711fe31da2a4
24/24under the fp32 contract (200/200 on GPU, 24/24 on CPU, incremental KVamber/FINDINGS.md:86 · 711fe31da2a4
84%teacher-forced encoder); bf16 decoders fail 84% of archives and fp16amber/FINDINGS.md:86 · 711fe31da2a4
28.5%bf16 decoders fail 84% of archives and fp16 28.5%, both inside the bandsamber/FINDINGS.md:86 · 711fe31da2a4
200/200- finding: cuda_fp32: 200/200 archives byte-exact,amber/FINDINGS.md:90 · 711fe31da2a4
0.33%coded depth (the fp32 analogue of F-8's 0.33% bf16amber/FINDINGS.md:90 · 711fe31da2a4
24/24mismatch is 0/49k). cpu_fp32: 24/24 (5,900 positions), i.e.amber/FINDINGS.md:90 · 711fe31da2a4
32/200kernel path from the encoder. cuda_bf16: 32/200 exact (fail 84%, bandamber/FINDINGS.md:90 · 711fe31da2a4
amber/FINDINGS.md:86 · 711fe31da2a4
F-12v3 binary archives are 130 B mean (4.3x smaller than the JSON v2 file, just under the predicted 5-7x)amber

Heading, as written v3 binary archives are 130 B mean (4.3x smaller than the JSON v2 file, just under the predicted 5-7x): 4.26 bits per generated token at L=245, of which 3.17 is the 97 B fixed header, 0.57 the draft-coded prompt, 0.45 the content; all 200 decode byte-exactly at fp32 with the prompt segment DECODED; the crossover vs zstd-19-concatenated is L~263 and the header stays >10% of the file until L~15,000

Sizes: mean file 130 B vs 551 B JSON v2 (ratio 0.231; the pre-registered 5-7x band is MISSED on the low side — the fixed floor of 97 B (pins 54 B incl. the 20 B snapshot, digests 24 B, frame+ints 19 B) is the file at this length, exactly as the protocol's own arithmetic said). amber/FINDINGS.md:102 · 711fe31da2a4

Regime
P1 content re-archived in v3 binary (src/amber/binformat.py: 5-byte frame, model id + 20 B snapshot + 8 B fingerprint, varint ints, three 8 B digests; ONE payload stream whose first segment codes prompt tokens 1..P-1 against the draft with the same {A,C,X} grammar, first token raw); encoder Llama-3.2-1B base cuda fp32 teacher-forced scoring every position (4,400 pos/s); … amber/FINDINGS.md:100 · 711fe31da2a4
The full regime line

P1 content re-archived in v3 binary (src/amber/binformat.py: 5-byte frame, model id + 20 B snapshot + 8 B fingerprint, varint ints, three 8 B digests; ONE payload stream whose first segment codes prompt tokens 1..P-1 against the draft with the same {A,C,X} grammar, first token raw); encoder Llama-3.2-1B base cuda fp32 teacher-forced scoring every position (4,400 pos/s); decode cuda fp32 incremental KV from the FIRST token (prompt reconstructed, not fed), all 200 archives; P5_PROTOCOL.md frozen with numeric bands before any archive was written

Source
  • results/p5_binformat.json
  • experiments/p5_binformat.py
  • experiments/P5_PROTOCOL.md
  • src/amber/binformat.py
  • src/amber/real/archive.py
  • tests/test_binformat.py
  • results/p5_decode.log
  • results/p5_report.log
amber/FINDINGS.md:101 · 711fe31da2a4
Date
2026-08-30 amber/FINDINGS.md:99 · 711fe31da2a4
4.26v2 file, just under the predicted 5-7x): 4.26 bits per generatedamber/FINDINGS.md:98 · 711fe31da2a4
3.17bits per generated token at L=245, of which 3.17 is the 97 B fixedamber/FINDINGS.md:98 · 711fe31da2a4
0.57of which 3.17 is the 97 B fixed header, 0.57 the draft-coded prompt,amber/FINDINGS.md:98 · 711fe31da2a4
0.45fixed header, 0.57 the draft-coded prompt, 0.45 the content; all 200amber/FINDINGS.md:98 · 711fe31da2a4
10%is L~263 and the header stays >10% of the file untilamber/FINDINGS.md:98 · 711fe31da2a4
0.231mean file 130 B vs 551 B JSON v2 (ratio 0.231; the pre-registeredamber/FINDINGS.md:102 · 711fe31da2a4
4.26own arithmetic said). Totals at L=245: 4.26 bits/generated tokenamber/FINDINGS.md:102 · 711fe31da2a4
4.0-4.6at L=245: 4.26 bits/generated token (band 4.0-4.6, in), = fixed 3.17 +amber/FINDINGS.md:102 · 711fe31da2a4
amber/FINDINGS.md:98 · 711fe31da2a4
F-13Session archives amortise the header exactly as the rate law says (4.26 -> 1.65 -> 1.45 bits/gen-token at k=1/8/16, 2.6x smaller than k separate files at k=8) and the shared-context mode LOSES on independently-generated turnsamber

Heading, as written Session archives amortise the header exactly as the rate law says (4.26 -> 1.65 -> 1.45 bits/gen-token at k=1/8/16, 2.6x smaller than k separate files at k=8) and the shared-context mode LOSES on independently-generated turns: conditioning the draft on history the target never saw raises content cost 16% at turn 2 and 39% by turn 8; reset-per-turn is the format, and it confines a desync to its turn

(1) Amortisation, reset mode: 2.79 / 2.03 / 1.65 / 1.45 bits/gen-token at k=2/4/8/16 (k=8 and 16 in their bands), with prompt (0.575) and content (0.440) terms flat as reset mode requires; header+table term 1.78 -> 0.44. amber/FINDINGS.md:110 · 711fe31da2a4

Regime
P1's 200 turns grouped in row order into k-turn sessions (k=2/4/8/16 -> 100/50/25/12 sessions; k=16 drops the last 8 rows); v4 session format (binformat.pack_session: one pin/digest set + per-turn table, ctx_mode byte); ctx_mode 0 assembled on CPU from P5's committed fp32 scores, ctx_mode 1 rescored on GPU fp32 teacher-forced in 512-token KV chunks; gamma, max_rank 64; decode cuda fp32 incremental KV (k=8: … amber/FINDINGS.md:108 · 711fe31da2a4
The full regime line

P1's 200 turns grouped in row order into k-turn sessions (k=2/4/8/16 -> 100/50/25/12 sessions; k=16 drops the last 8 rows); v4 session format (binformat.pack_session: one pin/digest set + per-turn table, ctx_mode byte); ctx_mode 0 assembled on CPU from P5's committed fp32 scores, ctx_mode 1 rescored on GPU fp32 teacher-forced in 512-token KV chunks; gamma, max_rank 64; decode cuda fp32 incremental KV (k=8: all 25 shared-mode sessions ~4.1k tokens each, first 8 reset-mode); P6_PROTOCOL.md frozen with bands before any session was written, including the structural-bias warning

Source
  • results/p6_sessions.json
  • experiments/p6_sessions.py
  • experiments/P6_PROTOCOL.md
  • src/amber/binformat.py
  • src/amber/real/archive.py
  • tests/test_sessions.py
  • results/p6_decode.log
  • results/p6_report.log
amber/FINDINGS.md:109 · 711fe31da2a4
Date
2026-08-30 amber/FINDINGS.md:107 · 711fe31da2a4
4.26the header exactly as the rate law says (4.26 -> 1.65 -> 1.45amber/FINDINGS.md:106 · 711fe31da2a4
1.65exactly as the rate law says (4.26 -> 1.65 -> 1.45 bits/gen-tokenamber/FINDINGS.md:106 · 711fe31da2a4
1.45as the rate law says (4.26 -> 1.65 -> 1.45 bits/gen-token atamber/FINDINGS.md:106 · 711fe31da2a4
16%the target never saw raises content cost 16% at turn 2 and 39% byamber/FINDINGS.md:106 · 711fe31da2a4
39%saw raises content cost 16% at turn 2 and 39% by turn 8;amber/FINDINGS.md:106 · 711fe31da2a4
2.79- finding: (1) Amortisation, reset mode: 2.79 / 2.03 / 1.65 / 1.45amber/FINDINGS.md:110 · 711fe31da2a4
2.03(1) Amortisation, reset mode: 2.79 / 2.03 / 1.65 / 1.45amber/FINDINGS.md:110 · 711fe31da2a4
1.65(1) Amortisation, reset mode: 2.79 / 2.03 / 1.65 / 1.45 bits/gen-tokenamber/FINDINGS.md:110 · 711fe31da2a4
amber/FINDINGS.md:106 · 711fe31da2a4
F-142026-10 errata pass: 12 of 16 checked claims corrected, 4 confirmed, 0 unrecomputable (3 decided only through …amber

Heading, as written 2026-10 errata pass: 12 of 16 checked claims corrected, 4 confirmed, 0 unrecomputable (3 decided only through logged amendments); the record is amended append-only from a pre-registered recompute of the committed evidence

one bullet per protocol row (claim as printed; recomputed value; status under the pinned rule, and where the protocol's amendment log says so, the status under the pinned rule before the amendment or in the claim's own regime). amber/FINDINGS.md:118 · 711fe31da2a4

Regime
recompute of printed claims from evidence committed at amber 070430b (tree 1c022e7), CPU only (torch blocked: true), python 3.11.9, numpy 2.3.5, zstandard 0.25.0, transformers 4.57.3 (E7's tokenizer only, offline); rules pre-registered in experiments/ERRATA_PROTOCOL.md (blob 22731ee5b8a0) with a dated amendment log; the review harvest's values (docs/ERRATA_PLAN.md) were scored predictions, never citations; … amber/FINDINGS.md:116 · 711fe31da2a4
The full regime line

recompute of printed claims from evidence committed at amber 070430b (tree 1c022e7), CPU only (torch blocked: true), python 3.11.9, numpy 2.3.5, zstandard 0.25.0, transformers 4.57.3 (E7's tokenizer only, offline); rules pre-registered in experiments/ERRATA_PROTOCOL.md (blob 22731ee5b8a0) with a dated amendment log; the review harvest's values (docs/ERRATA_PLAN.md) were scored predictions, never citations; every numeral in this entry is generated from results/errata_recompute.json by experiments/errata_text.py

Source
  • results/errata_recompute.json
  • experiments/ERRATA_PROTOCOL.md
  • experiments/errata_recompute.py
  • experiments/errata_lib.py
  • experiments/errata_text.py
  • tests/test_errata_recompute.py
  • tests/test_errata_text.py
  • results/errata_web/
  • results/errata_dryruns/
  • results/errata_critic/
  • docs/ERRATA_PLAN.md
amber/FINDINGS.md:117 · 711fe31da2a4
Date
2026-10-10 amber/FINDINGS.md:115 · 711fe31da2a4
12## F-14 — 2026-10 errata pass: 12 of 16 checked claimsamber/FINDINGS.md:114 · 711fe31da2a4
16## F-14 — 2026-10 errata pass: 12 of 16 checked claimsamber/FINDINGS.md:114 · 711fe31da2a4
4pass: 12 of 16 checked claims corrected, 4 confirmed, 0amber/FINDINGS.md:114 · 711fe31da2a4
016 checked claims corrected, 4 confirmed, 0 unrecomputable (3amber/FINDINGS.md:114 · 711fe31da2a4
3corrected, 4 confirmed, 0 unrecomputable (3 decided only throughamber/FINDINGS.md:114 · 711fe31da2a4
amber/FINDINGS.md:114 · 711fe31da2a4
F-152026-10 errata pass, second batch (rows E15, E16)amber

Heading, as written 2026-10 errata pass, second batch (rows E15, E16): 2 of 2 rows corrected under their pinned rules, 0 confirmed, 0 unrecomputable, both rules carrying a recorded defect, 7 of 8 claim-site statuses standing at their own scope; F-11's comparison of its failure rates with F-10's prediction crossed escape regimes while corrections stay more fragile than predicted, and F-14's "every one a near-tie" gets the threshold the record's own instrument uses while the top-1 claims it was read from stand

one bullet per row (claim as printed; recomputed value; status under the pinned rule, with the status in the claim's own regime, by symbol class, per decoder arm and per claim site where the protocol defines one). amber/FINDINGS.md:144 · 711fe31da2a4

Regime
recompute of printed claims from evidence committed at amber c3dc45e (tree f2e492c), CPU only (torch blocked: true), python 3.11.9, numpy 2.3.5; both rows pre-registered as the dated "second batch" entry of experiments/ERRATA_PROTOCOL.md, committed before the first code for them (the history shows that order; the order of writing is self-reported); … amber/FINDINGS.md:142 · 711fe31da2a4
The full regime line

recompute of printed claims from evidence committed at amber c3dc45e (tree f2e492c), CPU only (torch blocked: true), python 3.11.9, numpy 2.3.5; both rows pre-registered as the dated "second batch" entry of experiments/ERRATA_PROTOCOL.md, committed before the first code for them (the history shows that order; the order of writing is self-reported); the file's blob at this payload's head (e19df5a088eb) holds that entry, a post-dry-run entry (one informational field) and the post-critic entries that added the by-class, claim-scope and claim-site statuses in the brackets below and corrected the generated text; both rows were proposed by F-14's critic, whose values are this batch's scored predictions, never citations; every numeral in this entry is generated from results/errata_recompute_2.json by experiments/errata_text.py

Source
  • results/errata_recompute_2.json
  • experiments/ERRATA_PROTOCOL.md
  • experiments/errata_recompute.py
  • experiments/errata_lib.py
  • experiments/errata_text.py
  • tests/test_errata_recompute.py
  • tests/test_errata_text.py
  • results/errata_dryruns/dry_b2_pre.json
  • results/errata_dryruns/dry_b2.json
  • results/errata_critic/REPORT.md
amber/FINDINGS.md:143 · 711fe31da2a4
Date
2026-10-11 amber/FINDINGS.md:141 · 711fe31da2a4
2errata pass, second batch (rows E15, E16): 2 of 2 rows correctedamber/FINDINGS.md:140 · 711fe31da2a4
2pass, second batch (rows E15, E16): 2 of 2 rows corrected underamber/FINDINGS.md:140 · 711fe31da2a4
02 rows corrected under their pinned rules, 0 confirmed, 0amber/FINDINGS.md:140 · 711fe31da2a4
0under their pinned rules, 0 confirmed, 0 unrecomputable, bothamber/FINDINGS.md:140 · 711fe31da2a4
7both rules carrying a recorded defect, 7 of 8 claim-siteamber/FINDINGS.md:140 · 711fe31da2a4
8both rules carrying a recorded defect, 7 of 8 claim-site statusesamber/FINDINGS.md:140 · 711fe31da2a4
amber/FINDINGS.md:140 · 711fe31da2a4
F-162026-10 errata pass, third batch (rows E17, E18, E19)amberCONFIRMED

Heading, as written 2026-10 errata pass, third batch (rows E17, E18, E19): 2 of 3 rows corrected under their pinned rules, 1 confirmed, 0 unrecomputable, 3 of 14 claim-site statuses standing at their own scope; 6 of 7 printed "2tau" labels are 4 tau, the zstd baselines were strawmen at the per-archive scale, and the dumps' rank-1 desync rate is consistent with transfer to the archives once conditioned on the top-2 gap (post hoc at the primary binning; CONFIRMED at the others)

one bullet per row (claim as printed; recomputed value; status under the pinned rule, with the status per claim site, per count or under E15's rule where the protocol defines one). amber/FINDINGS.md:155 · 711fe31da2a4

In plain words
CONFIRMED
a recompute from committed evidence reproduced the value
Regime
recompute of printed claims from evidence committed at amber f492e68 (tree a800d10), CPU only (torch blocked: true), python 3.11.9, numpy 2.3.5, zstandard 0.25.0; the three rows pre-registered as the dated "third batch" entry of experiments/ERRATA_PROTOCOL.md, committed before the first code for them (the history shows that order; … amber/FINDINGS.md:153 · 711fe31da2a4
The full regime line

recompute of printed claims from evidence committed at amber f492e68 (tree a800d10), CPU only (torch blocked: true), python 3.11.9, numpy 2.3.5, zstandard 0.25.0; the three rows pre-registered as the dated "third batch" entry of experiments/ERRATA_PROTOCOL.md, committed before the first code for them (the history shows that order; the order of writing is self-reported), the file's blob at this payload's head being 4dd72905bc99; E18's predictions are the review harvest's, agent output scored after the run and never cited; E17's label ratios and E19's primary values (the second batch's critic's own post-hoc binning) had been seen before the entry, are recorded as such in the amendment log, and E19 is not scored; every numeral in this entry is generated from results/errata_recompute_3.json by experiments/errata_text.py

Source
  • results/errata_recompute_3.json
  • experiments/ERRATA_PROTOCOL.md
  • experiments/errata_recompute.py
  • experiments/errata_lib.py
  • experiments/errata_text.py
  • tests/test_errata_recompute.py
  • tests/test_errata_text.py
  • results/errata_dryruns/dry_b3.json
  • results/errata_dryruns/dry_b3b.json
  • results/p2_tinyllama_dumps.sha256
  • results/errata_critic/REPORT.md
amber/FINDINGS.md:154 · 711fe31da2a4
Date
2026-10-11 amber/FINDINGS.md:152 · 711fe31da2a4
2pass, third batch (rows E17, E18, E19): 2 of 3 rows correctedamber/FINDINGS.md:151 · 711fe31da2a4
3third batch (rows E17, E18, E19): 2 of 3 rows corrected underamber/FINDINGS.md:151 · 711fe31da2a4
13 rows corrected under their pinned rules, 1 confirmed, 0amber/FINDINGS.md:151 · 711fe31da2a4
0under their pinned rules, 1 confirmed, 0 unrecomputable, 3 of 14amber/FINDINGS.md:151 · 711fe31da2a4
3rules, 1 confirmed, 0 unrecomputable, 3 of 14 claim-siteamber/FINDINGS.md:151 · 711fe31da2a4
14rules, 1 confirmed, 0 unrecomputable, 3 of 14 claim-site statusesamber/FINDINGS.md:151 · 711fe31da2a4
6statuses standing at their own scope; 6 of 7 printed "2tau"amber/FINDINGS.md:151 · 711fe31da2a4
7statuses standing at their own scope; 6 of 7 printed "2tau" labelsamber/FINDINGS.md:151 · 711fe31da2a4
amber/FINDINGS.md:151 · 711fe31da2a4
F-1Pure-integer v0 pipeline round-trips bit-exactly, with expected ratio profilelockstep

30/30 unittest cases pass, including 80 seeded coder fuzz cases (varied alphabets 2–40 symbols, changing CDFs) and 6 conformance vectors. lockstep/trellis-docs/trellis-findings.md:28 · b9dff0132934

Context
run_002, this repo at v0.1.0, CPython 3.12.3 Linux x86_64 (envprobe digest in run row), stdlib only. lockstep/trellis-docs/trellis-findings.md:29 · b9dff0132934
Status
acted on (conformance/vectors_v0.json frozen; TP-1 unblocked). lockstep/trellis-docs/trellis-findings.md:31 · b9dff0132934
30/30**Observation**: 30/30 unittest cases pass,lockstep/trellis-docs/trellis-findings.md:28 · b9dff0132934
8030/30 unittest cases pass, including 80 seeded coder fuzz caseslockstep/trellis-docs/trellis-findings.md:28 · b9dff0132934
2seeded coder fuzz cases (varied alphabets 2–40 symbols, changinglockstep/trellis-docs/trellis-findings.md:28 · b9dff0132934
40seeded coder fuzz cases (varied alphabets 2–40 symbols, changing CDFs)lockstep/trellis-docs/trellis-findings.md:28 · b9dff0132934
6alphabets 2–40 symbols, changing CDFs) and 6 conformance vectors.lockstep/trellis-docs/trellis-findings.md:28 · b9dff0132934
4,394ratios: 1.87x on English markdown (4,394 B prose), 3.63x onlockstep/trellis-docs/trellis-findings.md:28 · b9dff0132934
32DRBG bytes. Pure-Python throughput ~32 KiB/s encode, ~30 KiB/slockstep/trellis-docs/trellis-findings.md:28 · b9dff0132934
0.52~30 KiB/s decode; 8 KiB random round-trip 0.52 s.lockstep/trellis-docs/trellis-findings.md:28 · b9dff0132934
lockstep/trellis-docs/trellis-findings.md:26 · b9dff0132934
F-2Integer probability grid absorbs sub-half-step drift; diverges past a step (the drift-sponge mechanism works)lockstep

On 200 seeded 128-symbol distributions at grid total 2^15 (step ≈ 3.06e-05 in probability space): alternating ±1e-11 perturbations produced identical quantized CDFs in 200/200 cases; ±1e-3 perturbations produced differing CDFs in 200/200 cases. lockstep/trellis-docs/trellis-findings.md:36 · b9dff0132934

Context
run_002, quantize_probs largest-remainder rule per specs/decode-profile-v0.md, tests_project/test_cdf.py drift test. lockstep/trellis-docs/trellis-findings.md:37 · b9dff0132934
Status
queued (E-001/E-002 measure real drift against this boundary; TP-2 builds the safeguard for what the grid can't absorb). lockstep/trellis-docs/trellis-findings.md:39 · b9dff0132934
200**Observation**: On 200 seeded 128-symbollockstep/trellis-docs/trellis-findings.md:36 · b9dff0132934
128**Observation**: On 200 seeded 128-symbol distributions atlockstep/trellis-docs/trellis-findings.md:36 · b9dff0132934
2128-symbol distributions at grid total 2^15 (step ≈ 3.06e-05 inlockstep/trellis-docs/trellis-findings.md:36 · b9dff0132934
3.06e-05distributions at grid total 2^15 (step ≈ 3.06e-05 in probability space):lockstep/trellis-docs/trellis-findings.md:36 · b9dff0132934
1e-11in probability space): alternating ±1e-11 perturbations producedlockstep/trellis-docs/trellis-findings.md:36 · b9dff0132934
200/200produced identical quantized CDFs in 200/200 cases; ±1e-3lockstep/trellis-docs/trellis-findings.md:36 · b9dff0132934
1e-3identical quantized CDFs in 200/200 cases; ±1e-3 perturbations producedlockstep/trellis-docs/trellis-findings.md:36 · b9dff0132934
200/200perturbations produced differing CDFs in 200/200 cases.lockstep/trellis-docs/trellis-findings.md:36 · b9dff0132934
lockstep/trellis-docs/trellis-findings.md:34 · b9dff0132934
F-3Untrusted-input review: coder is sound, container skin was not (now hardened)lockstep

Adversarial pass on the v0 container. lockstep/trellis-docs/trellis-findings.md:44 · b9dff0132934

Context
review session on v0.1.0; new module src/lockstep/manifest.py + tests_project/test_manifest.py added in response; suite 30 -> 42 tests, all green; all Trellis gates pass. lockstep/trellis-docs/trellis-findings.md:45 · b9dff0132934
Status
acted on (strict validate_manifest is now the single reader choke point; unpack/read_manifest take max_original_size; D-11 records the posture). lockstep/trellis-docs/trellis-findings.md:47 · b9dff0132934
636a sampled single-byte mutation sweep over a 636 B container producedlockstep/trellis-docs/trellis-findings.md:44 · b9dff0132934
0produced only clean ContainerError (0 silent successes, 0lockstep/trellis-docs/trellis-findings.md:44 · b9dff0132934
0clean ContainerError (0 silent successes, 0 leaked exceptions);lockstep/trellis-docs/trellis-findings.md:44 · b9dff0132934
0short-input round-trips (every length 0–64 x3 seeds, all 256lockstep/trellis-docs/trellis-findings.md:44 · b9dff0132934
64short-input round-trips (every length 0–64 x3 seeds, all 256lockstep/trellis-docs/trellis-findings.md:44 · b9dff0132934
256(every length 0–64 x3 seeds, all 256 single bytes,lockstep/trellis-docs/trellis-findings.md:44 · b9dff0132934
0/2555000x-repeat rescale path, alternating 0/255) all bit-exact;lockstep/trellis-docs/trellis-findings.md:44 · b9dff0132934
5bare KeyError/TypeError/AttributeError on 5 of 7 hand-craftedlockstep/trellis-docs/trellis-findings.md:44 · b9dff0132934
lockstep/trellis-docs/trellis-findings.md:42 · b9dff0132934
F-4First-cut cost model: under placeholder economics, LLM-archival loses to zstd -19 in every scenario, and …lockstep

Heading, as written First-cut cost model: under placeholder economics, LLM-archival loses to zstd -19 in every scenario, and compute cost dominates

tools/lockstep_costmodel.py, run on three built-in scenarios (compliance cold 1TB/84mo/2 reads; e-discovery warm 200GB/24mo/500 reads; research steward 5TB/120mo/50 reads) with placeholder-but-plausible lsp1 numbers (6x ratio, 0.5 MB/s pack, 2 MB/s decode, $2.50/GPU-hr): zstd -19 wins all three by 1–2 orders of magnitude, and pack/decode COMPUTE cost dominates so heavily there is no read-frequency crossover — e.g. lockstep/trellis-docs/trellis-findings.md:52 · b9dff0132934

The full first sentence

tools/lockstep_costmodel.py, run on three built-in scenarios (compliance cold 1TB/84mo/2 reads; e-discovery warm 200GB/24mo/500 reads; research steward 5TB/120mo/50 reads) with placeholder-but-plausible lsp1 numbers (6x ratio, 0.5 MB/s pack, 2 MB/s decode, $2.50/GPU-hr): zstd -19 wins all three by 1–2 orders of magnitude, and pack/decode COMPUTE cost dominates so heavily there is no read-frequency crossover — e.g. cold archive TCO $148 (zstd) vs $2,169 (lsp1), of which $1,389 is pack GPU-hours against only $56 of storage.

Context
review session; model inputs are PLACEHOLDERS (docstring says so); throughput must come from a real run (Blueprint B / tools/lockstep_bench.py) before any verdict is trusted. [withheld: Blueprint B not in this repository; see WITHHELD.md] lockstep/trellis-docs/trellis-findings.md:53 · b9dff0132934
Status
acted on (cost model committed as a standing tool; steers next work toward measuring real throughput and finding the survivable regime, not building more format features). lockstep/trellis-docs/trellis-findings.md:55 · b9dff0132934
0.5lsp1 numbers (6x ratio, 0.5 MB/s pack, 2 MB/slockstep/trellis-docs/trellis-findings.md:52 · b9dff0132934
2lsp1 numbers (6x ratio, 0.5 MB/s pack, 2 MB/s decode,lockstep/trellis-docs/trellis-findings.md:52 · b9dff0132934
2.50(6x ratio, 0.5 MB/s pack, 2 MB/s decode, $2.50/GPU-hr): zstd -19 winslockstep/trellis-docs/trellis-findings.md:52 · b9dff0132934
1$2.50/GPU-hr): zstd -19 wins all three by 1–2 orders of magnitude,lockstep/trellis-docs/trellis-findings.md:52 · b9dff0132934
2$2.50/GPU-hr): zstd -19 wins all three by 1–2 orders of magnitude,lockstep/trellis-docs/trellis-findings.md:52 · b9dff0132934
148crossover — e.g. cold archive TCO $148 (zstd) vs $2,169lockstep/trellis-docs/trellis-findings.md:52 · b9dff0132934
2,169— e.g. cold archive TCO $148 (zstd) vs $2,169 (lsp1), of which $1,389lockstep/trellis-docs/trellis-findings.md:52 · b9dff0132934
1,389TCO $148 (zstd) vs $2,169 (lsp1), of which $1,389 is pack GPU-hourslockstep/trellis-docs/trellis-findings.md:52 · b9dff0132934
lockstep/trellis-docs/trellis-findings.md:50 · b9dff0132934
F-5The coder's TOTAL_MAX=2^16 structurally forbids LLM vocabularieslockstep

Heading, as written The coder's TOTAL_MAX=2^16 structurally forbids LLM vocabularies; the constant is conservative, not load-bearing

`quantize_weights` requires total >= V and `Cdf` rejects total > TOTAL_MAX (2^16). lockstep/trellis-docs/trellis-findings.md:60 · b9dff0132934

Context
run_004 synthesis session; no new hardware; derived by crossing F-1/F-2 (measured on the 256-symbol byte alphabet at grid 2^15) against TP-1's plan to drive the same coder with an LLM vocabulary. lockstep/trellis-docs/trellis-findings.md:61 · b9dff0132934
Status
acted on (D-12 records the constants posture; TP-1/TP-2 marked BLOCKED; tools/lockstep_profile_advisor.py added so a command catches this class of error). lockstep/trellis-docs/trellis-findings.md:63 · b9dff0132934
2### F-5 — The coder's TOTAL_MAX=2^16 structurally forbidslockstep/trellis-docs/trellis-findings.md:58 · b9dff0132934
2>= V and `Cdf` rejects total > TOTAL_MAX (2^16). Measured: Llama-3lockstep/trellis-docs/trellis-findings.md:60 · b9dff0132934
128,256> TOTAL_MAX (2^16). Measured: Llama-3 (128,256), Qwen2-ish (151,936)lockstep/trellis-docs/trellis-findings.md:60 · b9dff0132934
2.0CDF at all; Mistral's 32,768 "fits" with **2.0 units/symbol and 50.0%lockstep/trellis-docs/trellis-findings.md:60 · b9dff0132934
50.0%32,768 "fits" with **2.0 units/symbol and 50.0% of probability masslockstep/trellis-docs/trellis-findings.md:60 · b9dff0132934
0.999by the mandatory >=1 floor**, costing **0.999 excess bits/tokenlockstep/trellis-docs/trellis-findings.md:60 · b9dff0132934
3.03costing **0.999 excess bits/token against a 3.03-bit distribution (~33%lockstep/trellis-docs/trellis-findings.md:60 · b9dff0132934
33%against a 3.03-bit distribution (~33% ratio loss)**.lockstep/trellis-docs/trellis-findings.md:60 · b9dff0132934
lockstep/trellis-docs/trellis-findings.md:58 · b9dff0132934
F-6The drift sponge does not scale: at LLM vocab, no grid both absorbs drift and preserves ratiolockstep

Re-ran F-2's exact methodology at LLM scale with PA-3's drift class (eps=1e-7). lockstep/trellis-docs/trellis-findings.md:68 · b9dff0132934

Context
run_004 synthesis session, peaky(tau=3) and zipf(1.0) synthetic distributions, TOTAL_MAX patched to 2^30 to make the test possible at all (see F-5). lockstep/trellis-docs/trellis-findings.md:69 · b9dff0132934
Status
acted on (D-13 moves the grid to logit space + top-k; OQ-4 closed; E-002/E-004 re-scoped; F-7 supplies the replacement mechanism). lockstep/trellis-docs/trellis-findings.md:71 · b9dff0132934
1e-7at LLM scale with PA-3's drift class (eps=1e-7). At V=32,768/T=2^15lockstep/trellis-docs/trellis-findings.md:68 · b9dff0132934
24/24the CDF was identical under drift 24/24 — but at **394% excesslockstep/trellis-docs/trellis-findings.md:68 · b9dff0132934
394%was identical under drift 24/24 — but at **394% excess entropy** (11.97lockstep/trellis-docs/trellis-findings.md:68 · b9dff0132934
11.9724/24 — but at **394% excess entropy** (11.97 excess bits vs 3.03lockstep/trellis-docs/trellis-findings.md:68 · b9dff0132934
3.03excess entropy** (11.97 excess bits vs 3.03 true), i.e. itlockstep/trellis-docs/trellis-findings.md:68 · b9dff0132934
1.5%preserving ratio (V=32,768 @ 2^20: 1.5% loss; V=128,256 @ 2^24:lockstep/trellis-docs/trellis-findings.md:68 · b9dff0132934
0.4%@ 2^20: 1.5% loss; V=128,256 @ 2^24: 0.4% loss) desynced **0/24lockstep/trellis-docs/trellis-findings.md:68 · b9dff0132934
0/24V=128,256 @ 2^24: 0.4% loss) desynced **0/24 identical — everylockstep/trellis-docs/trellis-findings.md:68 · b9dff0132934
lockstep/trellis-docs/trellis-findings.md:66 · b9dff0132934
F-7Logit-space quantization with top-k truncation restores sparse corrections; rate obeys k*eps/deltalockstep

Moving quantization from probability space (a global function of all V logits) to **logit space with top-k truncation** collapses the drift surface from V to k. lockstep/trellis-docs/trellis-findings.md:76 · b9dff0132934

Context
run_004 synthesis session, synthetic continuous logits. TWO earlier probes were discarded as invalid: (a) synthetic jitter on a decimal grid aligned with the coarse logit step, producing a spurious "coarser is worse" inversion; (b) a trial loop resampling perturbation signs but not base logits, producing a spurious 0% at every step. lockstep/trellis-docs/trellis-findings.md:77 · b9dff0132934
The full regime line

run_004 synthesis session, synthetic continuous logits. TWO earlier probes were discarded as invalid: (a) synthetic jitter on a decimal grid aligned with the coarse logit step, producing a spurious "coarser is worse" inversion; (b) a trial loop resampling perturbation signs but not base logits, producing a spurious 0% at every step. Both agreed with no theory, which is how they were caught.

Status
acted on (D-13; encoded in tools/lockstep_profile_advisor.py; E-006 will validate on real logits). lockstep/trellis-docs/trellis-findings.md:79 · b9dff0132934
1e-7logits per trial, V=4096, k=64, eps=1e-7): delta=1e-3 -> **0.67%lockstep/trellis-docs/trellis-findings.md:76 · b9dff0132934
1e-3per trial, V=4096, k=64, eps=1e-7): delta=1e-3 -> **0.67% measured vslockstep/trellis-docs/trellis-findings.md:76 · b9dff0132934
0.67%V=4096, k=64, eps=1e-7): delta=1e-3 -> **0.67% measured vs 0.64%lockstep/trellis-docs/trellis-findings.md:76 · b9dff0132934
0.64%delta=1e-3 -> **0.67% measured vs 0.64% predicted**; delta=1e-4lockstep/trellis-docs/trellis-findings.md:76 · b9dff0132934
1e-4measured vs 0.64% predicted**; delta=1e-4 -> 5.13% vs 6.40%;lockstep/trellis-docs/trellis-findings.md:76 · b9dff0132934
5.13%vs 0.64% predicted**; delta=1e-4 -> 5.13% vs 6.40%; delta=1e-2 ->lockstep/trellis-docs/trellis-findings.md:76 · b9dff0132934
6.40%0.64% predicted**; delta=1e-4 -> 5.13% vs 6.40%; delta=1e-2 -> 0/1500lockstep/trellis-docs/trellis-findings.md:76 · b9dff0132934
1e-2delta=1e-4 -> 5.13% vs 6.40%; delta=1e-2 -> 0/1500 vs 0.06%lockstep/trellis-docs/trellis-findings.md:76 · b9dff0132934
lockstep/trellis-docs/trellis-findings.md:74 · b9dff0132934
F-8Packing is 92% of TCO and needs a ~100x miraclelockstep

Heading, as written Packing is 92% of TCO and needs a ~100x miracle; the thesis survives only as a byproduct of inference already being run

F-4's cost model priced lsp1 decode at GPU rates, contradicting D-10 (CPU-first decode). lockstep/trellis-docs/trellis-findings.md:84 · b9dff0132934

Context
run_004 synthesis session; placeholder throughput/pricing inputs unchanged from F-4 (still unmeasured — see docs/08 and Blueprint B). [withheld: Blueprint B not in this repository; see WITHHELD.md] lockstep/trellis-docs/trellis-findings.md:85 · b9dff0132934
Status
queued (D-14 records the positioning bet as unvalidated; the next unit of work is a conversation with a prospective user, not more format code). lockstep/trellis-docs/trellis-findings.md:87 · b9dff0132934
92%### F-8 — Packing is 92% of TCO and needs alockstep/trellis-docs/trellis-findings.md:82 · b9dff0132934
1,502.67Re-priced (1 TB, 84 mo, 2 reads): lsp1 TCO $1,502.67 vs zstd -19 $148.41, oflockstep/trellis-docs/trellis-findings.md:84 · b9dff0132934
148.412 reads): lsp1 TCO $1,502.67 vs zstd -19 $148.41, of which **pack islockstep/trellis-docs/trellis-findings.md:84 · b9dff0132934
1,388.89vs zstd -19 $148.41, of which **pack is $1,388.89 = 92% of all cost**.lockstep/trellis-docs/trellis-findings.md:84 · b9dff0132934
92%-19 $148.41, of which **pack is $1,388.89 = 92% of all cost**.lockstep/trellis-docs/trellis-findings.md:84 · b9dff0132934
20.1= 92% of all cost**. Break-even requires **20.1 MB/s pack ≈ 5Mlockstep/trellis-docs/trellis-findings.md:84 · b9dff0132934
20%cost ~5x but loses at every text fraction (20% text -> $441; 5% textlockstep/trellis-docs/trellis-findings.md:84 · b9dff0132934
5%at every text fraction (20% text -> $441; 5% text -> $242). Ratio islockstep/trellis-docs/trellis-findings.md:84 · b9dff0132934
lockstep/trellis-docs/trellis-findings.md:82 · b9dff0132934
F-9The manifest claimed full model pinning while the reader ignored three semantic fieldslockstep

Starting from the reviewed/synthesis archives, independently forging `profile.model.alphabet`, `profile.model.increment`, or `profile.model.rescale_limit` left the archive acceptable and decodable; all three values were parsed but never compared with the selected model. lockstep/trellis-docs/trellis-findings.md:93 · b9dff0132934

Context
run_005 unification review of the three supplied repositories; each field was modified in an otherwise valid, re-framed container so payload and original hashes remained valid. lockstep/trellis-docs/trellis-findings.md:94 · b9dff0132934
Status
acted on (`validate_manifest` now requires the complete exact registered model profile, public profile tables are immutable, and adversarial regressions cover every field without changing any lsp0 vector). lockstep/trellis-docs/trellis-findings.md:96 · b9dff0132934
42with the selected model. The supplied 42-test product suitelockstep/trellis-docs/trellis-findings.md:93 · b9dff0132934
lockstep/trellis-docs/trellis-findings.md:91 · b9dff0132934
F-10A green visible suite masked a broken maintenance gate and an untested release surfacelockstep

The synthesis input passed 42/42 product tests but only 113/114 runnable Trellis tests (5 skipped): adding `lockstep-critic` raised the actual agent inventory to four while the expected count stayed three, contradicting run_003's recorded "all Trellis gates green." After unification, 82 product tests plus 39 pytest subtests pass, all 115 runnable Trellis tests pass with 4 precondition skips, the 768-case committed … lockstep/trellis-docs/trellis-findings.md:101 · b9dff0132934

The full first sentence

The synthesis input passed 42/42 product tests but only 113/114 runnable Trellis tests (5 skipped): adding `lockstep-critic` raised the actual agent inventory to four while the expected count stayed three, contradicting run_003's recorded "all Trellis gates green." After unification, 82 product tests plus 39 pytest subtests pass, all 115 runnable Trellis tests pass with 4 precondition skips, the 768-case committed mutation sweep and an extra 10,000-case malformed-container sweep leak no builtin exception, and a clean wheel runs all six packaged vectors outside the source tree.

Context
run_005 on CPython 3.13.5, Linux x86_64; synthesis used as the superset baseline, lsp0 conformance payloads held byte-identical. lockstep/trellis-docs/trellis-findings.md:102 · b9dff0132934
Status
acted on (agent inventory fixed, host test discovery generalized, package metadata/vector resources added, and the full repair gate plus clean-install check made part of the release procedure). lockstep/trellis-docs/trellis-findings.md:104 · b9dff0132934
42/42**Observation**: The synthesis input passed 42/42 product tests but onlylockstep/trellis-docs/trellis-findings.md:101 · b9dff0132934
113/114input passed 42/42 product tests but only 113/114 runnable Trellis testslockstep/trellis-docs/trellis-findings.md:101 · b9dff0132934
5but only 113/114 runnable Trellis tests (5 skipped): addinglockstep/trellis-docs/trellis-findings.md:101 · b9dff0132934
82Trellis gates green." After unification, 82 product tests plus 39lockstep/trellis-docs/trellis-findings.md:101 · b9dff0132934
39After unification, 82 product tests plus 39 pytest subtests pass,lockstep/trellis-docs/trellis-findings.md:101 · b9dff0132934
115tests plus 39 pytest subtests pass, all 115 runnable Trellis testslockstep/trellis-docs/trellis-findings.md:101 · b9dff0132934
4all 115 runnable Trellis tests pass with 4 precondition skips, thelockstep/trellis-docs/trellis-findings.md:101 · b9dff0132934
768tests pass with 4 precondition skips, the 768-case committed mutationlockstep/trellis-docs/trellis-findings.md:101 · b9dff0132934
lockstep/trellis-docs/trellis-findings.md:99 · b9dff0132934
F-11lsp0 conformance vectors reproduce byte-exactly on an independent OS/interpreter (Windows 11, CPython 3.x)lockstep

`lockstep self-check` reproduced all 6 lsp0 payload hashes and the full product suite passed 82/82 on a machine that shares nothing with prior validation: Windows 11 Pro 10.0.26200, CPython (MSVC build), envprobe digest `2d2a2e0c1d7795d5383d36bd22b9b51dfac7fde4171fa4f0db074f6de145bbbd`. lockstep/trellis-docs/trellis-findings.md:109 · b9dff0132934

Context
run_006, first use of this repo as a sibling dependency of the amber project (verified archives); installed editable from the lockstep-ultimate-repo.zip snapshot (sha256 4167766182777681d225d0c5c552341be31db8f027b11edb0707d5a6a8d6e409). No source changes. lockstep/trellis-docs/trellis-findings.md:110 · b9dff0132934
Status
noted (linked in trellis-runs.md). lockstep/trellis-docs/trellis-findings.md:112 · b9dff0132934
6`lockstep self-check` reproduced all 6 lsp0 payload hashes andlockstep/trellis-docs/trellis-findings.md:109 · b9dff0132934
82/82hashes and the full product suite passed 82/82 on a machine thatlockstep/trellis-docs/trellis-findings.md:109 · b9dff0132934
lockstep/trellis-docs/trellis-findings.md:107 · b9dff0132934
F-12First E-006 evidence (via amber, TinyLlama-1.1B, RTX 3080 Ti)lockstep

Heading, as written First E-006 evidence (via amber, TinyLlama-1.1B, RTX 3080 Ti): fp32 cross-device logit drift eps~3e-6 median / 1.3e-4 max; the k*eps/delta law holds; a delta=1e-3, k=64 grid crosses at 22% of positions even at fp32, and at 100% under bf16/fp16 — while token RANK order stayed exactly stable at fp32

Amber's Phase 2 census (amber F-5/F-6; `amber/results/p2_e006_grid_crossing.json`) dumped top-64 logits of TinyLlama-1.1B-Chat over 49,152 teacher-forced positions on CPU fp32 and on a 3080 Ti under fp32/bf16/fp16/int8/nf4. lockstep/trellis-docs/trellis-findings.md:117 · b9dff0132934

Context
run_007 (amber's capture, this repo's protocol); one CPU/GPU pair, one GPU architecture, one model; torch 2.6.0+cu124, cudnn autotune defaults; not the >=2-architecture census E-006 asks for. lockstep/trellis-docs/trellis-findings.md:118 · b9dff0132934
Status
queued (E-006 remains open for cross-architecture and burstiness; D-13's delta default should be revisited against this eps before TP-2 is unblocked). lockstep/trellis-docs/trellis-findings.md:120 · b9dff0132934
3e-63080 Ti): fp32 cross-device logit drift eps~3e-6 median / 1.3e-4 max;lockstep/trellis-docs/trellis-findings.md:115 · b9dff0132934
1.3e-4cross-device logit drift eps~3e-6 median / 1.3e-4 max; the k*eps/deltalockstep/trellis-docs/trellis-findings.md:115 · b9dff0132934
1e-3max; the k*eps/delta law holds; a delta=1e-3, k=64 grid crosses atlockstep/trellis-docs/trellis-findings.md:115 · b9dff0132934
22%holds; a delta=1e-3, k=64 grid crosses at 22% of positions even atlockstep/trellis-docs/trellis-findings.md:115 · b9dff0132934
100%at 22% of positions even at fp32, and at 100% under bf16/fp16 — whilelockstep/trellis-docs/trellis-findings.md:115 · b9dff0132934
3e-6CPU-vs-GPU fp32: |delta logit| p50 3e-6, p99 1.5e-5, max 1.3e-4lockstep/trellis-docs/trellis-findings.md:117 · b9dff0132934
1.5e-5fp32: |delta logit| p50 3e-6, p99 1.5e-5, max 1.3e-4 (logits oflockstep/trellis-docs/trellis-findings.md:117 · b9dff0132934
1.3e-4|delta logit| p50 3e-6, p99 1.5e-5, max 1.3e-4 (logits of magnitudelockstep/trellis-docs/trellis-findings.md:117 · b9dff0132934
lockstep/trellis-docs/trellis-findings.md:115 · b9dff0132934
F-13Second E-006 datapoint (Llama-3.2-1B-Instruct, V=128k)lockstep

Heading, as written Second E-006 datapoint (Llama-3.2-1B-Instruct, V=128k): fp32 cross-device eps p50 7e-6 / max 4.5e-4; delta=1e-3 grid crosses at 41% of positions at fp32 (law predicts 45%), delta=1e-2 at 5%; rank order stable to 1 flip in 49k

amber F-7 (`amber/results/p2_llama32_e006_grid_crossing.json`): same protocol as F-12 on the intended 1B Llama-3 draft with the 128,256-token vocabulary. lockstep/trellis-docs/trellis-findings.md:125 · b9dff0132934

Context
run_008; second model on the same box/stack as F-12 (one GPU architecture). Median eps is ~2x F-12's TinyLlama value; whether that tracks vocabulary size, depth, or logit scale is not separable from two models. lockstep/trellis-docs/trellis-findings.md:126 · b9dff0132934
Status
queued (D-13 delta default should move to ~1e-2 pending E-002's ratio cost; cross-architecture eps still unmeasured). lockstep/trellis-docs/trellis-findings.md:128 · b9dff0132934
7e-6V=128k): fp32 cross-device eps p50 7e-6 / max 4.5e-4;lockstep/trellis-docs/trellis-findings.md:123 · b9dff0132934
4.5e-4fp32 cross-device eps p50 7e-6 / max 4.5e-4; delta=1e-3 gridlockstep/trellis-docs/trellis-findings.md:123 · b9dff0132934
1e-3eps p50 7e-6 / max 4.5e-4; delta=1e-3 grid crosses at 41% oflockstep/trellis-docs/trellis-findings.md:123 · b9dff0132934
41%/ max 4.5e-4; delta=1e-3 grid crosses at 41% of positions at fp32lockstep/trellis-docs/trellis-findings.md:123 · b9dff0132934
45%at 41% of positions at fp32 (law predicts 45%), delta=1e-2 at 5%;lockstep/trellis-docs/trellis-findings.md:123 · b9dff0132934
1e-2positions at fp32 (law predicts 45%), delta=1e-2 at 5%; rank orderlockstep/trellis-docs/trellis-findings.md:123 · b9dff0132934
5%at fp32 (law predicts 45%), delta=1e-2 at 5%; rank order stable to 1lockstep/trellis-docs/trellis-findings.md:123 · b9dff0132934
7e-6CPU-vs-GPU fp32 |delta logit|: p50 7e-6, p99 2.9e-5, maxlockstep/trellis-docs/trellis-findings.md:125 · b9dff0132934
lockstep/trellis-docs/trellis-findings.md:123 · b9dff0132934
F-14First measured pack-side throughputs for the byproduct regime (amber Phase 1, 3080 Ti)lockstep

Heading, as written First measured pack-side throughputs for the byproduct regime (amber Phase 1, 3080 Ti): target generation 33 tok/s (3B bf16, batch 1, KV), draft rank scoring 9.2k tok/s (1B bf16); the archive itself costs 0.47 bits/token on target-generated text

amber F-8 (`amber/results/p1_divergence.json`): Llama-3.2-3B generated 48,958 tokens greedily at 32.8 tok/s; the 1B draft's teacher-forced replay (the archive's encode AND decode work) ran at 9,233 positions/s in bf16; the resulting amber v0 archive is 0.474 bits/token vs zstd -19 at 3.62 on the same UTF-8 text, tied with AC-under-draft (0.472). lockstep/trellis-docs/trellis-findings.md:133 · b9dff0132934

Context
run_009; consumer GPU, batch 1 generation (no batching), Python harness; measured, not projected (F-8's own rule). lockstep/trellis-docs/trellis-findings.md:134 · b9dff0132934
Status
queued (feed the measured 9.2k tok/s, batched, into tools/lockstep_costmodel.py as the pack input and re-run the three scenarios — the OQ-13 conversation now has a number to bring). lockstep/trellis-docs/trellis-findings.md:136 · b9dff0132934
1for the byproduct regime (amber Phase 1, 3080 Ti): targetlockstep/trellis-docs/trellis-findings.md:131 · b9dff0132934
3080for the byproduct regime (amber Phase 1, 3080 Ti): target generationlockstep/trellis-docs/trellis-findings.md:131 · b9dff0132934
9.2k(3B bf16, batch 1, KV), draft rank scoring 9.2k tok/s (1B bf16); thelockstep/trellis-docs/trellis-findings.md:131 · b9dff0132934
0.47tok/s (1B bf16); the archive itself costs 0.47 bits/token onlockstep/trellis-docs/trellis-findings.md:131 · b9dff0132934
32.8generated 48,958 tokens greedily at 32.8 tok/s; the 1B draft'slockstep/trellis-docs/trellis-findings.md:133 · b9dff0132934
0.474in bf16; the resulting amber v0 archive is 0.474 bits/token vs zstd -19lockstep/trellis-docs/trellis-findings.md:133 · b9dff0132934
3.62archive is 0.474 bits/token vs zstd -19 at 3.62 on the same UTF-8 text,lockstep/trellis-docs/trellis-findings.md:133 · b9dff0132934
0.472same UTF-8 text, tied with AC-under-draft (0.472).lockstep/trellis-docs/trellis-findings.md:133 · b9dff0132934
lockstep/trellis-docs/trellis-findings.md:131 · b9dff0132934
F-15Rank-ORDER stability has a depthlockstep

Heading, as written Rank-ORDER stability has a depth: on the Llama-3.2-1B dumps an fp32 decoder reproduces an fp32 encoder's token order exactly to rank 64 (0 desyncs / 49k positions), while bf16 mis-orders 3% of rank-2 and 9-26% of rank-3..8 tokens and a top-1 near-tie escape cannot buy that back; lockstep.rac is now amber's production coder (value-free contexts) and gains only 1% over Elias-gamma there

amber F-10 (`amber/results/p2b_escape_robustness.json`, run_008's dumps re-read at top-256): symbol streams built under cpu_fp32 and resolved under each GPU arm. lockstep/trellis-docs/trellis-findings.md:141 · b9dff0132934

Context
run_010; one model, one box, teacher-forced text (deeper ranks than target-generated text; per-rank rates transfer, the mix does not); cross-architecture still pending. amber wrote its own small-alphabet adaptive model because `models.AdaptiveByteModel` hard-codes a 256-symbol alphabet. lockstep/trellis-docs/trellis-findings.md:142 · b9dff0132934
Status
verified (amber tests/test_rac_codec.py exercise rac end-to-end incl. corruption; p2b counts are deterministic re-reads of committed dumps). lockstep/trellis-docs/trellis-findings.md:144 · b9dff0132934
3%/ 49k positions), while bf16 mis-orders 3% of rank-2 and 9-26% oflockstep/trellis-docs/trellis-findings.md:139 · b9dff0132934
9-26%while bf16 mis-orders 3% of rank-2 and 9-26% of rank-3..8 tokens andlockstep/trellis-docs/trellis-findings.md:139 · b9dff0132934
1%coder (value-free contexts) and gains only 1% over Elias-gamma therelockstep/trellis-docs/trellis-findings.md:139 · b9dff0132934
1.3e-5arm. cuda_fp32_b1: 1 desync at theta=0 (gap 1.3e-5), 0 at any escapelockstep/trellis-docs/trellis-findings.md:141 · b9dff0132934
5.1%margin, across all ranks to 64. cuda_fp16: 5.1% of rank-2..64lockstep/trellis-docs/trellis-findings.md:141 · b9dff0132934
0.4%rank-2..64 corrections mis-resolve (rank 2 0.4%, rank 8 3.7%, rankslockstep/trellis-docs/trellis-findings.md:141 · b9dff0132934
3.7%mis-resolve (rank 2 0.4%, rank 8 3.7%, ranks 17-64 14%).lockstep/trellis-docs/trellis-findings.md:141 · b9dff0132934
17-64(rank 2 0.4%, rank 8 3.7%, ranks 17-64 14%). cuda_bf16 (b1 orlockstep/trellis-docs/trellis-findings.md:141 · b9dff0132934
lockstep/trellis-docs/trellis-findings.md:139 · b9dff0132934
F-16The fp32 replay contract holds end-to-end on real archiveslockstep

Heading, as written The fp32 replay contract holds end-to-end on real archives: 200/200 GPU and 24/24 CPU draft-only decodes byte-exact across device AND kernel path (teacher-forced encode vs incremental KV decode, 0 desyncs / 55k positions); bf16 decoders fail 84% of archives, fp16 28.5%

amber F-11 (`amber/results/p4_real_archive.json`): Llama-3.2-1B base archives of 48,958 target-generated tokens (200 prompts), encoded on cuda fp32 with one teacher-forced forward per prompt, decoded draft-only with incremental KV replay. lockstep/trellis-docs/trellis-findings.md:149 · b9dff0132934

Context
run_011; one GPU architecture; Python batch-1 decode at 26 tok/s (overhead-bound, fp32 not slower than bf16); encode 991 pos/s. lockstep/trellis-docs/trellis-findings.md:150 · b9dff0132934
Status
verified (amber tests/test_hf_archive.py cover the pure core; the run is committed evidence). lockstep/trellis-docs/trellis-findings.md:152 · b9dff0132934
200/200contract holds end-to-end on real archives: 200/200 GPU and 24/24 CPUlockstep/trellis-docs/trellis-findings.md:147 · b9dff0132934
24/24on real archives: 200/200 GPU and 24/24 CPU draft-only decodeslockstep/trellis-docs/trellis-findings.md:147 · b9dff0132934
84%/ 55k positions); bf16 decoders fail 84% of archives, fp16 28.5%lockstep/trellis-docs/trellis-findings.md:147 · b9dff0132934
28.5%bf16 decoders fail 84% of archives, fp16 28.5%lockstep/trellis-docs/trellis-findings.md:147 · b9dff0132934
200/200with incremental KV replay. cuda_fp32 200/200 exact; cpu_fp32 24/24;lockstep/trellis-docs/trellis-findings.md:149 · b9dff0132934
24/24replay. cuda_fp32 200/200 exact; cpu_fp32 24/24; cuda_bf16 32/200lockstep/trellis-docs/trellis-findings.md:149 · b9dff0132934
32/200200/200 exact; cpu_fp32 24/24; cuda_bf16 32/200 (first desync p50 atlockstep/trellis-docs/trellis-findings.md:149 · b9dff0132934
143/20028, rank-3 27, rank-4 16); cuda_fp16 143/200. Every top-1 desync satlockstep/trellis-docs/trellis-findings.md:149 · b9dff0132934
lockstep/trellis-docs/trellis-findings.md:147 · b9dff0132934
F-17Erratum to F-14: the pack rate is 0.0437 MB/s, not ~5 MB/s; the units slipped by 114.5x, and the …lockstepCONFIRMEDCORRECTED

Heading, as written Erratum to F-14: the pack rate is 0.0437 MB/s, not ~5 MB/s; the units slipped by 114.5x, and the target-generation rate was an unweighted shard mean

amber's errata pass recomputed F-14's inputs from committed evidence (amber 0f6d2f0 (`results/errata_recompute.json`, blob 551a1ccd868b; F-14 there), rows E10 and E5). lockstep/trellis-docs/trellis-findings.md:161 · b9dff0132934

In plain words
CONFIRMED
a recompute from committed evidence reproduced the value
CORRECTED
a later dated erratum corrected a number or a sentence in this entry
Context
no new measurement; the same P1 shards F-14 cited (amber `results/p1/draft/meta.json`, `results/p1/generate/meta.json`, `results/p1_divergence.json`), recomputed CPU-only under a pre-registered protocol (amber `experiments/ERRATA_PROTOCOL.md`) and independently re-derived by a critic (amber `results/errata_critic/REPORT.md`, recompute discrepancy 0). lockstep/trellis-docs/trellis-findings.md:162 · b9dff0132934
The full regime line

no new measurement; the same P1 shards F-14 cited (amber `results/p1/draft/meta.json`, `results/p1/generate/meta.json`, `results/p1_divergence.json`), recomputed CPU-only under a pre-registered protocol (amber `experiments/ERRATA_PROTOCOL.md`) and independently re-derived by a critic (amber `results/errata_critic/REPORT.md`, recompute discrepancy 0). MB = 10^6 bytes, as in F-8.

Status
noted (append-only erratum; F-14's text stands as written). Withdrawn: "the economics are no longer two orders of magnitude off"; "within ~4x of its 20 MB/s break-even"; "10-20x above F-8's 0.5 MB/s placeholder". lockstep/trellis-docs/trellis-findings.md:164 · b9dff0132934
0.0437### F-17 — Erratum to F-14: the pack rate is 0.0437 MB/s, not ~5 MB/s; thelockstep/trellis-docs/trellis-findings.md:159 · b9dff0132934
4.7277(token-weighted over 25 shards) at 4.7277 bytes/token of decodedlockstep/trellis-docs/trellis-findings.md:161 · b9dff0132934
4.7277The pack rate is their product: 9,233 x 4.7277 = 43,651 B/s = 0.0437lockstep/trellis-docs/trellis-findings.md:161 · b9dff0132934
0.0437product: 9,233 x 4.7277 = 43,651 B/s = 0.0437 MB/s. Against F-14'slockstep/trellis-docs/trellis-findings.md:161 · b9dff0132934
0.5"~5 MB/s" that is 114.5x lower; against the 0.5 MB/s pack placeholderlockstep/trellis-docs/trellis-findings.md:161 · b9dff0132934
20.1called it 10-20x higher); against F-8's 20.1 MB/s break-even 460xlockstep/trellis-docs/trellis-findings.md:161 · b9dff0132934
32.8"within ~4x"). Row E5 (CORRECTED): F-14's 32.8 tok/s target generationlockstep/trellis-docs/trellis-findings.md:161 · b9dff0132934
25.0of 25 per-shard rates; token-weighted it is 25.0 tok/s (48958 tokens inlockstep/trellis-docs/trellis-findings.md:161 · b9dff0132934
lockstep/trellis-docs/trellis-findings.md:159 · b9dff0132934
F-18Status of F-15 and F-16 after amber's errata passlockstepCONFIRMEDCORRECTED

Heading, as written Status of F-15 and F-16 after amber's errata pass: the fp32 "0 desyncs" is the escape-on count (1 at the shipped theta 0; 1 at batch 8 at every theta), the per-rank table was the theta 0.465 row, "991 pos/s" is a stall-dominated aggregate, and "0 desyncs / 55k" stands

from amber 0f6d2f0 (`results/errata_recompute.json`, blob 551a1ccd868b; F-14 there): row E8a: cpu_fp32 -> cuda_fp32_b1 has 1 desync in 49,056 positions at theta 0 (the shipped default) and 0 with the escape on, so F-15's headline "0 desyncs / 49k positions" is the escape-on count (its Observation already said so). lockstep/trellis-docs/trellis-findings.md:169 · b9dff0132934

In plain words
CONFIRMED
a recompute from committed evidence reproduced the value
CORRECTED
a later dated erratum corrected a number or a sentence in this entry
Context
no new measurement; amber's committed P2b payload, P4 decode shards and encode metadata, recomputed under amber `experiments/ERRATA_PROTOCOL.md` with its dated amendment log (E9's pinned rule checked the table at the wrong theta and is recorded there as a protocol defect) and re-derived by the critic (discrepancy 0). lockstep/trellis-docs/trellis-findings.md:170 · b9dff0132934
Status
noted (append-only; F-15 and F-16 stand as written with this entry as their qualifier). lockstep/trellis-docs/trellis-findings.md:172 · b9dff0132934
0.465theta), the per-rank table was the theta 0.465 row, "991 pos/s" is alockstep/trellis-docs/trellis-findings.md:167 · b9dff0132934
2.96e-05at decoder rank 5; the swapped pair sits 2.96e-05 apart), which answerslockstep/trellis-docs/trellis-findings.md:169 · b9dff0132934
0.465amber F-10's, and that table is the theta 0.465 / m 2 row of the P2blockstep/trellis-docs/trellis-findings.md:169 · b9dff0132934
6.0%on), unstated in F-10; at theta 0 rank 2 is 6.0% (bf16) and 0.7% (fp16)lockstep/trellis-docs/trellis-findings.md:169 · b9dff0132934
0.7%F-10; at theta 0 rank 2 is 6.0% (bf16) and 0.7% (fp16) against thelockstep/trellis-docs/trellis-findings.md:169 · b9dff0132934
2.7%(bf16) and 0.7% (fp16) against the printed 2.7% and 0.4%; the otherlockstep/trellis-docs/trellis-findings.md:169 · b9dff0132934
0.4%0.7% (fp16) against the printed 2.7% and 0.4%; the other printedlockstep/trellis-docs/trellis-findings.md:169 · b9dff0132934
49.4as the aggregate (991 pos/s over 200 rows, 49.4 s) but 17 rows belowlockstep/trellis-docs/trellis-findings.md:169 · b9dff0132934
lockstep/trellis-docs/trellis-findings.md:167 · b9dff0132934
F-19Closing finding: the lsp1 profile was never built; this research is closedlockstep

no lsp1 profile exists in the source tree at this repository's last research commit, as README.md states ("not implemented"). lockstep/trellis-docs/trellis-findings.md:177 · b9dff0132934

Context
TP-7 (`task-packs/done/TP-7_errata_and_archive.md`), 2026-10-10; starting state 2d7d079; amber's errata pass at 0f6d2f0. Search command: `git grep -n -i lsp1 2d7d079 -- src`; output: 0 lines (empty). lockstep/trellis-docs/trellis-findings.md:178 · b9dff0132934
Status
acted on (README banner; task packs TP-1 to TP-6 archived; no further findings will be appended). lockstep/trellis-docs/trellis-findings.md:180 · b9dff0132934
92%measured economics rest on F-8 (packing is 92% of TCO) and F-17 (thelockstep/trellis-docs/trellis-findings.md:177 · b9dff0132934
0.0437TCO) and F-17 (the byproduct pack rate is 0.0437 MB/s, 460x short oflockstep/trellis-docs/trellis-findings.md:177 · b9dff0132934
lockstep/trellis-docs/trellis-findings.md:175 · b9dff0132934
F-20Post-archive pointer: amber's errata batches 2 and 3 correct three sentences this log restated (F-15's …lockstepCORRECTED

Heading, as written Post-archive pointer: amber's errata batches 2 and 3 correct three sentences this log restated (F-15's "per-rank rates transfer", F-16's prediction comparison, F-18's "every one a near-tie"); this repository is harvested into sideband and parked

amber's errata batches 2 and 3 (amber 69b437c, tag `errata-batch-3-landed`) landed after TP-7 cited batch 1 (0f6d2f0) and correct three sentences this log restated from amber. lockstep/trellis-docs/trellis-findings.md:185 · b9dff0132934

In plain words
CORRECTED
a later dated erratum corrected a number or a sentence in this entry
Context
no new measurement. Values read by key from amber's committed payloads at 69b437cba973 by the sideband harvest generator and re-derived by an independent critic; sideband `docs/DECODE_CONTRACTS.md` carries them with a Sources table of file, blob and key. lockstep/trellis-docs/trellis-findings.md:186 · b9dff0132934
The full regime line

no new measurement. Values read by key from amber's committed payloads at 69b437cba973 by the sideband harvest generator and re-derived by an independent critic; sideband `docs/DECODE_CONTRACTS.md` carries them with a Sources table of file, blob and key. The sites in this repository: F-15's Context, F-16's Observation, F-18's Why it matters, and the F-15 row of the Linked findings table in `trellis-docs/trellis-runs.md`.

Status
acted on (TP-8). Harvested into sideband at e74cbd78e17d7a9065ea7cf7ab6e526174bc47b2 (branch `espalier-trial`): …
The full status line

acted on (TP-8). Harvested into sideband at e74cbd78e17d7a9065ea7cf7ab6e526174bc47b2 (branch `espalier-trial`): `docs/DECODE_CONTRACTS.md` (the value contract's grid crossings beside amber's order-contract mismatches, every number generated from pinned commits), a dated addition to `docs/METHOD.md` section 6 (F-14's 114.5x pack-rate error), and dated notes in sideband's OQ-7, BLUEPRINT T-A, GLOSSARY and ECOSYSTEM saying that this repository's byproduct economics were never measured and its runtime never built. F-19's "no further findings will be appended" is amended by this entry, which is the last. Repository parked 2026-10-11.

lockstep/trellis-docs/trellis-findings.md:188 · b9dff0132934
63.8first desyncs (bf16 28 observed against 63.8 expected, z -4.5; fp16lockstep/trellis-docs/trellis-findings.md:185 · b9dff0132934
-4.5(bf16 28 observed against 63.8 expected, z -4.5; fp16 9 against 35.2, zlockstep/trellis-docs/trellis-findings.md:185 · b9dff0132934
35.263.8 expected, z -4.5; fp16 9 against 35.2, z -4.4; row E15),lockstep/trellis-docs/trellis-findings.md:185 · b9dff0132934
-4.4expected, z -4.5; fp16 9 against 35.2, z -4.4; row E15), whilelockstep/trellis-docs/trellis-findings.md:185 · b9dff0132934
33.3on the encoder's top-2 gap it fits (33.3 against 28, z -0.9;lockstep/trellis-docs/trellis-findings.md:185 · b9dff0132934
-0.9top-2 gap it fits (33.3 against 28, z -0.9; 15.0 against 9, zlockstep/trellis-docs/trellis-findings.md:185 · b9dff0132934
15.0top-2 gap it fits (33.3 against 28, z -0.9; 15.0 against 9, z -1.6),lockstep/trellis-docs/trellis-findings.md:185 · b9dff0132934
-1.6(33.3 against 28, z -0.9; 15.0 against 9, z -1.6), post hoc at thelockstep/trellis-docs/trellis-findings.md:185 · b9dff0132934
lockstep/trellis-docs/trellis-findings.md:183 · b9dff0132934
F-21Erratum to F-2: the integer grid does not absorb drift "up to about half a grid step"; F-2's own generator …lockstep

Heading, as written Erratum to F-2: the integer grid does not absorb drift "up to about half a grid step"; F-2's own generator flips the quantized CDF at a median of 0.02 of a half step, and nothing survives a quarter step

The image audit of 2026-10-18 (docs/audits/2026-10-18-image-audit/README.md, section B item 1; finders R-8 and S1-C-1) re-ran F-2's generator (tests_project/test_cdf.py: 200 cases, V=128, total 2^15) at perturbations between F-2's two endpoints: 165 of 200 quantized CDFs identical at 1e-7, 98 at 3e-7, 10 at 1e-6 and 0 at 2e-6, against a half step of 1.53e-5; the median case flips at 0.02 of a half step and none survives a quarter step; … lockstep/trellis-docs/trellis-findings.md:193 · b9dff0132934

The full first sentence

The image audit of 2026-10-18 (docs/audits/2026-10-18-image-audit/README.md, section B item 1; finders R-8 and S1-C-1) re-ran F-2's generator (tests_project/test_cdf.py: 200 cases, V=128, total 2^15) at perturbations between F-2's two endpoints: 165 of 200 quantized CDFs identical at 1e-7, 98 at 3e-7, 10 at 1e-6 and 0 at 2e-6, against a half step of 1.53e-5; the median case flips at 0.02 of a half step and none survives a quarter step; the boundary scales as step/V.

Context
no new instrument; the audit's rerun of the committed generator. The confirming test, tests_project/test_cdf.py line 45, perturbs by 1e-11 ("orders below half a grid step") and so cannot see the boundary. lockstep/trellis-docs/trellis-findings.md:194 · b9dff0132934
Status
noted (append-only erratum; F-2's text stands as written; dated pointers beside docs/01 item 2 and docs/04 item 1). Withdrawn as boundary statements: "absorbs sub-half-step drift"; "a measurable boundary near half a grid step". lockstep/trellis-docs/trellis-findings.md:196 · b9dff0132934
0.02flips the quantized CDF at a median of 0.02 of a half step, andlockstep/trellis-docs/trellis-findings.md:191 · b9dff0132934
1-10-18-image-audit/README.md, section B item 1; finders R-8 andlockstep/trellis-docs/trellis-findings.md:193 · b9dff0132934
1e-7165 of 200 quantized CDFs identical at 1e-7, 98 at 3e-7, 10 at 1e-6lockstep/trellis-docs/trellis-findings.md:193 · b9dff0132934
3e-7200 quantized CDFs identical at 1e-7, 98 at 3e-7, 10 at 1e-6 and 0 atlockstep/trellis-docs/trellis-findings.md:193 · b9dff0132934
1e-6CDFs identical at 1e-7, 98 at 3e-7, 10 at 1e-6 and 0 at 2e-6, againstlockstep/trellis-docs/trellis-findings.md:193 · b9dff0132934
2e-6at 1e-7, 98 at 3e-7, 10 at 1e-6 and 0 at 2e-6, against a half step oflockstep/trellis-docs/trellis-findings.md:193 · b9dff0132934
1.53e-51e-6 and 0 at 2e-6, against a half step of 1.53e-5; the median case flipslockstep/trellis-docs/trellis-findings.md:193 · b9dff0132934
0.02step of 1.53e-5; the median case flips at 0.02 of a half step and nonelockstep/trellis-docs/trellis-findings.md:193 · b9dff0132934
lockstep/trellis-docs/trellis-findings.md:191 · b9dff0132934

The three parts

One repository, three records, published together as they stood at their own pinned commits. Each card gives the question the record asked, where it ended, and how many findings it carries.

sideband · e3d7ea76b93d

sideband

Can the memory-access address stream a weight-serving device sees for free steer a speculative-decoding drafter enough to matter?

Existence ABSORBED by the full free baseline at the pinned cells, with the L3 cells left open (S1 L3 NOT-SHOWN, S3 L3 SPLIT); conversion not shown in two registered attempts at S1; the reconstruction target at S1 proved to be free information. The table below carries the record's own words per layer and source

The experiment above, plus two side results. README.md:187 · c7c4f9562e72

30 findings sideband/FINDINGS.md:68–833 · d147e8c0cd6b README.md:108 · c7c4f9562e72 sideband/WITHHELD.md:3 · b01fa29dbd2b
amber · 0b061b54d92c

amber

Archives of model-generated text coded as a draft model's rank corrections, decoded with the draft alone, byte-exact under a measured contract

A codec with a measured rate, a decode contract shown sufficient on one box, and an errata pass that corrected 16 of the 21 claims it checked

An archive of target-generated text as the draft model's rank corrections, encoded from the draft's teacher-forced logits over the target's text and decoded by the draft alone, byte-exact under an fp32 contract; the draft codes both ends and the target only generates (F-14 row E1, which corrected the record's first line). README.md:249 · c7c4f9562e72

16 findings amber/FINDINGS.md:6–151 · 711fe31da2a4 README.md:109 · c7c4f9562e72 amber/WITHHELD.md:3 · fd35447b9bd0
lockstep · a84bef4968dc

lockstep

A pure-integer arithmetic coder with conformance vectors, and a proposed model-pinned archive profile

The byte-model coder and its vectors stand; the model-backed profile was never built and the economics behind it did not survive measurement

A pure-integer 32-bit arithmetic coder with bit-exact conformance vectors for its byte-model profile lsp0, a container that pins coder constants and every model parameter, and typed failure categories. README.md:334 · c7c4f9562e72

21 findings lockstep/trellis-docs/trellis-findings.md:26–191 · b9dff0132934 README.md:110 · c7c4f9562e72 lockstep/WITHHELD.md:3 · c57265724b55

Errata and dated notes

When a registered text turned out to be wrong, the record did not rewrite it. It appended a dated erratum, note or amendment beside the text. This timeline is those lines: the 70 lines of the three findings logs that carry a date and one of the words erratum, errata, dated note, amendment, correction, superseded or struck (47 in sideband, 18 in amber, 5 in lockstep), quoted as written, newest month first, each opening from a short preview to the full line. A line is filed under the first date it carries; on a row of sideband's findings table that is the finding's own date, and the note's date comes later in the row.

October 202641 entries

  1. 2026-10-18sidebandF-23dated note

    DATED NOTE 2026-10-18 (prior art for the identity; the cold read of that date, docs/PUBLIC_CUT_PLAN.md 15.8): …

    The full line, as written
    - DATED NOTE 2026-10-18 (prior art for the identity; the cold read of that date, docs/PUBLIC_CUT_PLAN.md 15.8): the identity this erratum rests on, that the final-norm state is the lm_head input, so the full logits are its image under the lm_head matrix and, that matrix having full column rank (the check this erratum's payload records), the state follows from them, is the setting of the published softmax-bottleneck work: Finlayson, Ren and Swayamdipta, "Logits of API-Protected LLMs Leak Proprietary Information" (arXiv 2403.09539, 2024), which shows the logits confined to a linear subspace of dimension equal to the hidden size, and Carlini et al., "Stealing Part of a Production Language Model" (arXiv 2403.06634, 2024), which recovers the embedding projection layer up to symmetries from API access; Finlayson, Ren and Swayamdipta do not claim recovery of the state, and Carlini et al. recover the projection rather than a state; the state's recovery follows from the identity here. Both fetched on 2026-10-18 and added to docs/PRIOR_ART.md (its 2026-10-18 additions). The record already concedes the order of events in docs/IDEAS.md line 247 ("textbook prior art (Finlayson 2024, Carlini 2024) that nobody in the lane applied": Stage R ran and Stage A was frozen before the identity was applied). This note places that concession against the literature; the erratum's text stands as written. [withheld: docs/IDEAS.md, docs/PUBLIC_CUT_PLAN.md not in this repository; see WITHHELD.md]
    sideband/FINDINGS.md:269 · d147e8c0cd6b
  2. 2026-10-18sidebandF-24dated note

    DATED NOTE 2026-10-18 (the first free baseline and the routing literature; the cold read of that date, docs/PUBLIC_CUT_PLAN.md 15.8): the free baseline of F-15 and F-17 left token ids out "as non-ordinal" (tools/run_probes.py line 196, a frozen instrument).

    The full line, as written
    - DATED NOTE 2026-10-18 (the first free baseline and the routing literature; the cold read of that date, docs/PUBLIC_CUT_PLAN.md 15.8): the free baseline of F-15 and F-17 left token ids out "as non-ordinal" (tools/run_probes.py line 196, a frozen instrument). At S3, where the stream is OLMoE's router address stream, the absorption this finding measured once hashed predicted-token ids (TOK) and the z-scored h_a block entered the baseline together is consistent with two published analyses of MoE routing: OpenMoE (Xue et al., arXiv 2402.01739, 2024) reports routing decisions that depend mainly on the routed input token's ID with little regard to context (TOK holds predicted ids, so of the two only OLMoE's output-id sentence bears on TOK directly), and OLMoE (Muennighoff et al., arXiv 2409.02060, 2024), the paper of the base model behind S3's checkpoint (OLMoE-1B-7B-0125-Instruct, docs/METHOD.md line 269; the paper's section 5.4 analyses OLMoE-1B-7B), reports vocabulary specialization of its experts, later layers specializing on predicted output token IDs. Consistent with, not predicted by: the re-test added TOK and the standardized hz block in one arm, so the share of the absorption due to token ids alone was not isolated, and at S1 the stream is top-32-of-512 weight-slot occupancy at the down_proj inputs of a dense Llama (docs/GLOSSARY.md, the telemetry-sources entry), where the routing literature does not apply. Both papers fetched on 2026-10-18 and added to docs/PRIOR_ART.md. Categorical features are routinely one-hot or hash encoded; the exclusion was a design choice the re-test revisited, and the labels stand as written. The probing-practice references added the same day (Pimentel et al. 2020; Hewitt and Liang 2019; Xu et al. 2020) bear on the readers: untuned, scored on top-1 accuracy, and at F-15 and F-17 on raw inputs (this finding's reader reads z-scored h_a); the K-2 word is scoped to the registered reader throughout, and those references name what a stronger reader would be. [withheld: docs/PUBLIC_CUT_PLAN.md not in this repository; see WITHHELD.md]
    sideband/FINDINGS.md:278 · d147e8c0cd6b
  3. 2026-10-18amberF-11dated note

    dated note (2026-10-18): what the fp32 contract pins is the dtype alone. The archive header's model block carries id, snapshot and dtype (src/amber/real/archive.py line 39, MODEL_FIELDS) and the capture arm adds device, quantization and batch (src/amber/real/capture.py lines 89-95, Arm); …

    The full line, as written
    - dated note (2026-10-18): what the fp32 contract pins is the dtype alone. The archive header's model block carries id, snapshot and dtype (src/amber/real/archive.py line 39, MODEL_FIELDS) and the capture arm adds device, quantization and batch (src/amber/real/capture.py lines 89-95, Arm); matmul precision, the attention implementation, the torch and transformers versions and the GPU architecture are recorded in the capture's meta.json (capture.py's docstring) and in no archive, so a decoder on another stack is bound by nothing in the file. The bf16 and fp16 failure rates in this title are properties of this PyTorch stack at these settings, not of the formats: ts_zip evaluates its model in BF16 and states that the result "does not depend on the exact GPU or CPU model nor on the number of configured threads" (bellard.org/ts_zip, fetched 2026-10-18; docs/PRIOR_ART.md, 2026-10-18 additions). The counts stand; the dtype pin is a measurement of this stack, not a contract that binds another.
    amber/FINDINGS.md:95 · 711fe31da2a4
  4. 2026-10-18amberF-11dated note

    dated note (2026-10-18, second, after the critic's recompute of the same day): the note above misstates the scope of the two rates.

    The full line, as written
    - dated note (2026-10-18, second, after the critic's recompute of the same day): the note above misstates the scope of the two rates. P4's archives were encoded on cuda fp32 (experiments/P4_PROTOCOL.md line 18) and the bf16 and fp16 decoders read those fp32-encoded archives, so the 84% and 28.5% are cross-dtype figures, a property of the dtype difference between encoder and decoder under this stack; the phrase "not of the formats" is withdrawn. What ts_zip's page bears on is a same-dtype contract, bf16 at both ends with deterministic kernels, which P4 did not measure. The header's model block binds a decoder to nothing beyond id, snapshot and dtype; "bound by nothing in the file" above reads with that qualification.
    amber/FINDINGS.md:96 · 711fe31da2a4
  5. 2026-10-18lockstepF-8dated note

    **Dated note (2026-10-18)**: the figures in this Observation rest on the cost model's placeholder inputs, the 2 MB/s decode rate among them (docs/audits/2026-10-18-image-audit/README.md, item B3).

    The full line, as written
    **Dated note (2026-10-18)**: the figures in this Observation rest on the cost model's placeholder inputs, the 2 MB/s decode rate among them (docs/audits/2026-10-18-image-audit/README.md, item B3). With any decode rate the record measured in that placeholder's place, lsp1 costs more than zstd -19 even with packing free, so the free-packing branch ("wins by 23%") does not survive measured inputs; the verdict that the standalone thesis loses survives every input set the audit's refuters tried, and the reason given for it, the 92% share, does not. The Observation stands as written.
    lockstep/trellis-docs/trellis-findings.md:89 · b9dff0132934
  6. 2026-10-18lockstepF-16dated note

    **Dated note (2026-10-18)**: the bf16 and fp16 failure rates in this title are a property of amber's PyTorch inference stack at its settings (one GPU architecture, batch 1, the Hugging Face model code), not of the numeric formats.

    The full line, as written
    **Dated note (2026-10-18)**: the bf16 and fp16 failure rates in this title are a property of amber's PyTorch inference stack at its settings (one GPU architecture, batch 1, the Hugging Face model code), not of the numeric formats. ts_zip evaluates its model in BF16 and states that the result "does not depend on the exact GPU or CPU model nor on the number of configured threads" (bellard.org/ts_zip, fetched 2026-10-18; quoted under PA-4 in docs/02-prior-art.md). F-16's counts stand; the sentence "bf16 decoders fail 84% of archives" reads with that scope, and the cross-GPU open question above is the one ts_zip's engineering answers for its own model.
    lockstep/trellis-docs/trellis-findings.md:155 · b9dff0132934
  7. 2026-10-18lockstepF-16dated note

    **Dated note (2026-10-18, second, after the critic's recompute of the same day)**: the note above misstates the scope of the two rates.

    The full line, as written
    **Dated note (2026-10-18, second, after the critic's recompute of the same day)**: the note above misstates the scope of the two rates. amber's P4 archives were encoded on cuda fp32 (amber experiments/P4_PROTOCOL.md line 18) and the bf16 and fp16 decoders read those fp32-encoded archives, so the 84% and 28.5% are cross-dtype figures, a property of the dtype difference between encoder and decoder under that stack; "not of the numeric formats" is withdrawn. What ts_zip's page bears on is a same-dtype contract, bf16 at both ends with deterministic kernels, which P4 did not measure and which the cross-GPU open question above would need.
    lockstep/trellis-docs/trellis-findings.md:157 · b9dff0132934
  8. 2026-10-16sidebandF-3dated note

    regime: synthetic concept-HMM v0, default WorldConfig; gamma=4; 16k positions; oracle policies Dated clause 2026-10-16: steering mechanism argmax, run_simulation's parameter default; the script passes no steer_mech, so D-10's posterior default did not apply (numbers unaffected; …

    The full line, as written
    - regime: synthetic concept-HMM v0, default WorldConfig; gamma=4; 16k positions; oracle policies Dated clause 2026-10-16: steering mechanism argmax, run_simulation's parameter default; the script passes no steer_mech, so D-10's posterior default did not apply (numbers unaffected; DECISIONS.md D-10, dated note).
    sideband/FINDINGS.md:90 · d147e8c0cd6b
  9. 2026-10-16sidebandF-4dated note

    regime: synthetic concept-HMM v0, default WorldConfig except rho; disagreement-gated beta=2.5; gamma=6; 14k positions; analytic per-position acceptance (unconditional, D-9) Dated clause 2026-10-16: steering mechanism argmax, run_simulation's parameter default; …

    The full line, as written
    - regime: synthetic concept-HMM v0, default WorldConfig except rho; disagreement-gated beta=2.5; gamma=6; 14k positions; analytic per-position acceptance (unconditional, D-9) Dated clause 2026-10-16: steering mechanism argmax, run_simulation's parameter default; the script passes no steer_mech, so D-10's posterior default did not apply (numbers unaffected; DECISIONS.md D-10, dated note); one seed (200) per curve.
    sideband/FINDINGS.md:97 · d147e8c0cd6b
  10. 2026-10-16sidebandF-4erratum

    DATED ERRATUM 2026-10-16 (image audit 2026-10-15, section B item 5; docs/audits/2026-10-15-image-audit/f4_seeds.py reruns e04's protocol at seeds 200 to 209, the seed-200 run reproducing the committed payload to 1e-12): "(net harmful at every offset)" at rho=0.6 is not supported at offset 1.

    The full line, as written
    DATED ERRATUM 2026-10-16 (image audit 2026-10-15, section B item 5; docs/audits/2026-10-15-image-audit/f4_seeds.py reruns e04's protocol at seeds 200 to 209, the seed-200 run reproducing the committed payload to 1e-12): "(net harmful at every offset)" at rho=0.6 is not supported at offset 1. The stale-minus-vanilla analytic acceptance at offset 1 reads -0.0039, +0.0001, -0.0015, -0.0012, -0.0003, -0.0022, +0.0047, +0.0010, +0.0029, +0.0057 across the ten seeds, mean +0.0005 (SE 0.0010), negative on 5 of 10, and the committed seed-200 value is the most negative of the ten. Offsets 2 to 6 are negative on 10 of 10 (means -0.0121 / -0.0194 / -0.0228 / -0.0259 / -0.0267), and offset 1 at rho 0.75 and 0.9 is positive on 10 of 10 (means +0.0097 / +0.0174), so the decay curve and the rule of thumb stand; the offset-1 sign at rho 0.6 is seed noise. Every e04 curve is one seed (200); the mechanism is argmax (the dated clause on the regime line).
    sideband/FINDINGS.md:101 · d147e8c0cd6b
  11. 2026-10-16sidebandF-5dated note

    regime: synthetic concept-HMM v0, hardened telemetry (support_gain=0.5, support_noise=0.5, background_noise=0.25, top_n=16), n_slots swept; disagreement-gated beta=2.5; gamma=4; 12k positions; sidecar = softmax regression on slot multi-hot Dated clause 2026-10-16: steering mechanism argmax, run_simulation's parameter default; …

    The full line, as written
    - regime: synthetic concept-HMM v0, hardened telemetry (support_gain=0.5, support_noise=0.5, background_noise=0.25, top_n=16), n_slots swept; disagreement-gated beta=2.5; gamma=4; 12k positions; sidecar = softmax regression on slot multi-hot Dated clause 2026-10-16: steering mechanism argmax, run_simulation's parameter default; the script passes no steer_mech, so D-10's posterior default did not apply (numbers unaffected; DECISIONS.md D-10, dated note).
    sideband/FINDINGS.md:105 · d147e8c0cd6b
  12. 2026-10-16sidebandF-6dated note
    - regime: synthetic concept-HMM v0 — structural property of the world, not a tuning outcome. Dated clause 2026-10-16: the e03 numbers quoted here ran at the argmax mechanism, run_simulation's parameter default (DECISIONS.md D-10, dated note).
    sideband/FINDINGS.md:112 · d147e8c0cd6b
  13. 2026-10-16sidebandF-9erratum

    DATED ERRATUM 2026-10-16 (image audit 2026-10-15, section B item 5; docs/audits/2026-10-15-image-audit/e07_ledger.py reruns e07's three arms with the bonus token priced, the correction bits reproducing the committed payload to four places): …

    The full line, as written
    DATED ERRATUM 2026-10-16 (image audit 2026-10-15, section B item 5; docs/audits/2026-10-15-image-audit/e07_ledger.py reruns e07's three arms with the bonus token priced, the correction bits reproducing the committed payload to four places): the v0 ledger omits the bonus token the verifier emits when a whole block is accepted, a term that is not arm-invariant: 0.0628 / 0.0654 / 0.0684 bonus tokens per emitted token on vanilla / free-causal / fresh-hint at 5.54 / 5.52 / 5.57 bits each, 0.348 / 0.361 / 0.381 unpriced bits per emitted token. With it priced the savings read 0.0862 (free-causal) and 0.1291 (fresh hint) bits per emitted token against the committed 0.0997 and 0.1621, so the committed figures are 16% and 26% high and the TOTAL break-even budget of 0.162 bits/token is overstated by about a quarter; the dense hint's loss reads about 26:1 on the fuller ledger (20:1 as committed). Direction and shape stand; the accept/reject indicator stream stays unpriced as caveated (its sign depends on the coding convention).
    sideband/FINDINGS.md:137 · d147e8c0cd6b
  14. 2026-10-16sidebandF-21dated note

    DATED NOTE 2026-10-16 (image audit 2026-10-15, section B item 3; docs/audits/2026-10-15-image-audit/replay_c1.py and replay_c1_seed6.py with their JSON outputs; NON-CRITERION, no label moves): the frozen Stage-C heads ran live with NO support clip.

    The full line, as written
    - DATED NOTE 2026-10-16 (image audit 2026-10-15, section B item 3; docs/audits/2026-10-15-image-audit/replay_c1.py and replay_c1_seed6.py with their JSON outputs; NON-CRITERION, no label moves): the frozen Stage-C heads ran live with NO support clip. HeadScorer applies only the [0, 1] output clip, and the exported weights carry TRAIN moments but no TRAIN support, so the standardization pathology F-22 (iii) records for Stage R, which Stage I was shielded from by predict_clipped, reached the live scorer here; this entry said nothing about it. The audit's read-only replay of the logged decisions from the gitignored per-prompt byproduct arrays (not reproducible from the committed tree) matches every logged stop_pos / extend decision of the seed-3 verdict arms: 17,185 of 17,185 consulted blocks on arm 6 (C3, free features) and 17,616 of 17,616 on arm 7 (C4, telemetry). Standardized inputs on the consulted decision path reach |z| 17,531 (arm 6) and 122,318 (arm 7); 111 / 120 consulted blocks exceed |z| 100 at offset 1. The honest exposure measure is the counterfactual TRAIN-support clip (the DR-1 form): it changes 41 decisions on arm 6 (0.24% of consulted blocks) and 41 on arm 7 (0.23%), in both directions; the voided seed-2 arms read 41 / 30 of 16,970 / 17,257, and the OQ-20 seed-6 arms 42 / 47 of 17,157 / 17,602. A first-order estimate of the direct tokens/call effect is at most 0.0119 / 0.0116 (41 decisions times at most five drafted tokens, a stop at position 1 against an extension to six, over the consulted calls; the rest of a changed prompt's trajectory is not priced; probable, not rerun), far inside the C4-vs-C3 interval -0.081846 [-0.132484, -0.035567] and the vs-vanilla straddle, so CONVERSION-MODE NULL stands as read. Rule recorded in docs/LESSONS.md: a scorer that runs live carries its TRAIN support clip, pinned at the freeze, not only in diagnostics.
    sideband/FINDINGS.md:248 · d147e8c0cd6b
  15. 2026-10-16sidebandF-22dated note

    WHAT DID NOT SURVIVE — THE MAGNITUDES (critic B2): Standardizer.from_moments floors only exactly-zero sd, and rare mass512 buckets carry TRAIN sd down to ~1e-7.

    The full line, as written
    - WHAT DID NOT SURVIVE — THE MAGNITUDES (critic B2): Standardizer.from_moments floors only exactly-zero sd, and rare mass512 buckets carry TRAIN sd down to ~1e-7. Two oq15 TEST prompts (orig 140, 37) put mass into such buckets and produce standardized inputs of 26,662 and 14,082; ridge at 1e5 cannot guard an unbounded input, and EVERY trained arm lands 4-8x worse than the mean predictor on those clusters (per-cluster MSE triv/tel/free/noise 0.969 / 5.760 / 8.265 / 8.084 and 0.975 / 3.495 / 4.704 / 4.529; 9.4% + 5.4% of pooled TEL SSE). Consequences, critic-recomputed: (i) TEL-vs-TRIVIAL pooled +0.10701 [-0.06983, +0.22088] — the honest deployable margin against a perfectly-shrunk baseline is NOT CI-clear at the registered pooled estimand (k3 +0.2163 [+0.2037, +0.2287] clear; oq15 -0.0002 [-0.3557, +0.2238]); TEL beats trivial in 81/84 clusters with median per-cluster gain +0.2243 (k3) / +0.2251 (oq15), the three losers being exactly the pathology prompts (140, 37, 206). (ii) The registered deltas are pathology-INFLATED, not deflated: the critic's UNREGISTERED drop-two diagnostic (quoted only as such) gives delta_noise +0.1327 [+0.1278, +0.1373], delta_free +0.1569 [+0.1504, +0.1629], TEL-vs-TRIVIAL +0.2162 [+0.2050, +0.2263], label still PASS — label robust, magnitudes ~a quarter pathology. (iii) Stage I (F-20) shared the standardization and was shielded by predict_clipped's [0,1] clip; the unbounded target removed the shield. (dated note 2026-10-16: the live Stage-C heads of F-21 ran with no support clip either; F-21's dated note of the same date records the replayed exposure, 41 changed decisions per arm under a counterfactual TRAIN-support clip) 4,063 TEST cells exceed 20 sd over the 657 free-feature columns (6,212 over the full TEL design) — the pathology is widespread at lower magnitude. The DR-1 clip is an OUT-OF-SUPPORT guard, not an sd floor: only 425 of those 6,212 cells lie outside TRAIN support, the clip changes 446 TEST cells in total, and ~5,800 in-support >20-sd cells remain (post-clip max |z| 204.1, ~sqrt(n_train)).
    sideband/FINDINGS.md:256 · d147e8c0cd6b
  16. 2026-10-16sidebandF-24dated note

    finding: MECHANICAL LABELS: S1 ABSORBED (horizon OFF3-ABSORBED); S3 ABSORBED — both at the registered reader at the registered budget, both preconditions met on every harvest (reader validity: min gate-cell accuracy minus floor +0.0468 / +0.0393 at S1, +0.0272 / +0.0207 at S3; Mf gate CI upper +0.0475 / +0.0432 / +0.0298 / +0.0241; …

    The full line, as written
    - finding: MECHANICAL LABELS: S1 ABSORBED (horizon OFF3-ABSORBED); S3 ABSORBED — both at the registered reader at the registered budget, both preconditions met on every harvest (reader validity: min gate-cell accuracy minus floor +0.0468 / +0.0393 at S1, +0.0272 / +0.0207 at S3; Mf gate CI upper +0.0475 / +0.0432 / +0.0298 / +0.0241; leak rider: shuffled fplus 0.5733 / 0.5681 / 0.5792 / 0.5851 against committed shuffled (a) 0.5889 / 0.5871 / 0.5799 / 0.5858, shuffled fplus_b 0.5740 / 0.5706 / 0.5787 / 0.5845, both legs pass on all four). (i) THE GATE, S1: D at L1/off2/mlp256 +0.0002 [-0.0012, +0.0015] (per seed +0.0017 / -0.0006 / -0.0007) primary, +0.0012 [-0.0003, +0.0028] (+0.0016 / +0.0013 / +0.0007) replication — fails all three conditions on the primary and two on the replication; Mt beside it +0.0034 [+0.0021, +0.0048] (+0.0036 / +0.0051 / +0.0016) and +0.0033 [+0.0012, +0.0052] (+0.0043 / +0.0015 / +0.0040) — CI-clear positive on both harvests, below the +0.005 materiality floor at the point on both, its replication upper +0.0052 a hair above the floor; Mn, the shift twin's own gain, +0.0033 [+0.0018, +0.0049] and +0.0020 [-0.0000, +0.0041]. Linear off2 D (reported) +0.0009 [-0.0006, +0.0024] / +0.0012 [-0.0002, +0.0026]; off3 mlp256 D -0.0004 [-0.0011, +0.0002] / +0.0005 [-0.0012, +0.0022] with Mt +0.0010 [+0.0002, +0.0020] / +0.0024 [+0.0002, +0.0047] — every one of the six S1 D rows has its CI95 upper below +0.005 (max +0.0028), the registered ABSORBED condition, and the offset-3 horizon reads OFF3-ABSORBED by the same rule. S3: D +0.0003 [-0.0013, +0.0020] (-0.0003 / +0.0019 / -0.0007) primary, +0.0001 [-0.0019, +0.0021] (+0.0010 / +0.0002 / -0.0009) replication; Mt +0.0033 [+0.0013, +0.0054] (+0.0038 / +0.0023 / +0.0038) and +0.0013 [-0.0008, +0.0035] (+0.0026 / +0.0023 / -0.0010) — CI-clear on the primary only, below the floor on both; Mn +0.0030 [+0.0014, +0.0046] and +0.0012 [-0.0004, +0.0026]. Linear off2 D +0.0015 [+0.0001, +0.0029] (+0.0004 / +0.0002 / +0.0039) primary — CI-POSITIVE-IMMATERIAL, and 78% of its point is the shift twin's own linear COST (Mn linear -0.0012 [-0.0027, +0.0003]; Mt linear only +0.0003 [-0.0009, +0.0016]) — and +0.0010 [-0.0003, +0.0024] replication (dated note 2026-10-16: the primary's descriptive qualifier is seed-sensitive, reading [-0.0013, +0.0042] / [-0.0009, +0.0039] / [-0.0008, +0.0047] under the three seed-inclusive approximations of docs/audits/2026-10-15-image-audit/seed_var_out.json, so CI-POSITIVE is lost under all three and IMMATERIAL stands; and the percentile cluster bootstrap covers about 0.939 at 30 clusters, docs/audits/2026-10-15-image-audit/coverage_out.json, so its lower bound +0.0001 sits inside the bootstrap's own shortfall while its upper +0.0029 reads about +0.0030 under a t-for-z correction; no label moves); the four S3 D uppers are +0.0029 / +0.0020 / +0.0024 / +0.0021. So the licensed sentence at both sources is "no MATERIAL (+0.005) equal-dims margin at the registered reader", never "no margin": aligned telemetry still adds about +0.003 over the full free baseline at the mlp256 gate (Mt), CI-clear on three of the four harvests, and nothing beyond the equal-dims twin (D). (ii) THE FREE BASELINE MOVED: Mf = fplus - (a) at the gate +0.0369 [+0.0266, +0.0475] / +0.0330 [+0.0223, +0.0432] (S1) and +0.0253 [+0.0208, +0.0298] / +0.0176 [+0.0101, +0.0241] (S3), larger than F-15's and F-17's entire telemetry margins over (a) at that cell; and the full free baseline ALONE beats the committed (a)+(b) telemetry arm at mlp256 — Mc +0.0112 [+0.0061, +0.0165] / +0.0130 [+0.0059, +0.0204] (S1), +0.0052 [+0.0019, +0.0084] (S3 primary; the S3 replication ties, +0.0009 [-0.0046, +0.0056]); at linear Mc is -0.0000 [-0.0079, +0.0090] / +0.0037 [-0.0035, +0.0123] (S1) and +0.0006 [-0.0026, +0.0038] / -0.0042 [-0.0088, -0.0000] (S3) — on the S3 replication the registered reader given 4753 free dims scores below the committed 1681-dim (a)+(b) linear probe, the reader-scope caveat with teeth. Gate-cell accuracies (fplus / fplus_b / fplus_n per seed; floors 0.5914 / 0.5897 / 0.5803 / 0.5865): S1 primary 0.6454 0.6381 0.6475 / 0.6490 0.6433 0.6490 / 0.6473 0.6438 0.6497; S1 replication 0.6290 0.6403 0.6347 / 0.6332 0.6418 0.6387 / 0.6316 0.6405 0.6380; S3 primary 0.6083 0.6094 0.6075 / 0.6121 0.6117 0.6113 / 0.6124 0.6099 0.6119; S3 replication 0.6072 0.6113 0.6084 / 0.6098 0.6137 0.6074 / 0.6088 0.6135 0.6083. The committed (a) mlp256 cells at those cells were 0.6062-0.6071 / 0.5967-0.5990 (S1) and 0.5824-0.5844 / 0.5914 (S3). (iii) THE TWIN WAS NOT INERT, AND THE CORRECTED TWIN READ: the S1 critic recompute found what the registration had not anticipated — the shift twin reads a position 17 tokens ahead of the anchor, whose forward pass has consumed the label token (the offset-2 label sits at a+1) for 93.7% (S1) / 93.5% (S3) of anchors, and it gains CI-clear over the base on both primary gates (Mn above), about as much as the aligned block, so D = Mt - Mn sits at zero by subtraction; F-16 had named that channel only as one that "could only make M2n an UNDERestimate", which is true under SURVIVES and runs toward the null under a kill. At S3 the dependence is confined to the primary harvest (the replication's Mn and both its Mt uppers sit below the floor); at S1 the ABSORBED-versus-NOT-SHOWN call on the replication rides on D specifically (with an inert twin D would equal Mt, whose upper +0.0052 is above the floor). At S1 the corrected twin read TWIN-ROBUST — D_r +0.0002 [-0.0009, +0.0014] primary (seeds +0.0013 / -0.0016 / +0.0008), +0.0013 [-0.0002, +0.0028] replication (seeds +0.0029 / +0.0009 / -0.0001); Mn_r +0.0032 [+0.0016, +0.0050] / +0.0020 [+0.0002, +0.0037]; T +0.0000 [-0.0012, +0.0013] / +0.0000 [-0.0016, +0.0019] (TWIN-INFORMATION-NOT-SHOWN) — so the registered null does not depend on the shift twin and the consequences act exactly as pinned. DATED NOTE 2026-10-16 (seed sensitivity; docs/audits/2026-10-15-image-audit/seed_var.py, as in F-25's note of the same date; NON-CRITERION, no label moves): the S1 corrected twin's CI-clear Mn_r reads [-0.0006, +0.0071] / [-0.0003, +0.0068] / [+0.0001, +0.0071] on the primary (per seed +0.0023 / +0.0067 / +0.0007) and [-0.0007, +0.0047] / [+0.0003, +0.0037] / [-0.0008, +0.0052] on the replication (per seed +0.0013 / +0.0006 / +0.0041) under the three seed-inclusive approximations (normal combination / interaction-corrected method of moments / prompt x seed pigeonhole bootstrap), so CI-clear is seed-sensitive at S1, lost under two of the three on each harvest with all six per-seed points positive; the S3 primary Mn_r stays CI-clear under all three. TWIN-ROBUST and every D of this entry stand: the largest D upper under the three approximations is +0.0042 / +0.0039 / +0.0047, below the +0.005 floor. On the primary and replication the corrected twin itself gains CI-clear over the base while T's CI includes zero, so the label-consumption mechanism is WITHDRAWN for both harvests: the shift twin's gain is matched by a prompt-independent block, which under the pinned reading means a same-marginal sparse block of these dims yields about Mt at this reader — the strongest form of the null, under which Mt stops being evidence of telemetry information (the registered reading of one construction, not a measurement of every block); D_r remains the matched comparison. At S3 the corrected twin read TWIN-ROBUST — D_r +0.0000 [-0.0014, +0.0014] primary (seeds +0.0010 / +0.0001 / -0.0009), +0.0008 [-0.0014, +0.0029] replication (seeds +0.0019 / +0.0009 / -0.0006); Mn_r +0.0033 [+0.0015, +0.0050] / +0.0005 [-0.0012, +0.0022]; T -0.0003 [-0.0021, +0.0015] / +0.0007 [-0.0005, +0.0018] (TWIN-INFORMATION-NOT-SHOWN) — so the registered null does not depend on the shift twin and the consequences act exactly as pinned. On the primary the corrected twin itself gains CI-clear over the base while T's CI includes zero, so the label-consumption mechanism is WITHDRAWN for that harvest: the shift twin's gain is matched by a prompt-independent block, which under the pinned reading means a same-marginal sparse block of these dims yields about Mt at this reader — the strongest form of the null, under which Mt stops being evidence of telemetry information (the registered reading of one construction, not a measurement of every block); D_r remains the matched comparison. On the replication the mechanism is neither withdrawn nor supported (Mn_r and T straddle zero). NET AT THIS READER, a gate-cell (mlp256) statement: the three sparse blocks measured — the aligned telemetry, the same row's telemetry 17 positions ahead, and a different prompt's telemetry at the same anchor index — add +0.0005..+0.0034 over the full free baseline (Mt / Mn / Mn_r: CI-clear for all three on both primaries, on the S1 replication for the aligned block and, by +0.0002, the prompt-level twin, and for none on the S3 replication), and the aligned block carries no MATERIAL margin beyond either twin at this power (D and D_r points +0.0000..+0.0013 with every upper at or below +0.0029; the S1 replication's D and D_r lowers are -0.0003 and -0.0002, so "nothing" is not licensed); a helpful twin lowers D_r below Mt by the identity D_r = Mt - Mn_r, which is the matched comparison, not dilution. The pinned reading extends the one prompt-independent construction to any same-marginal sparse block of these dims — a registered reading, not a measurement. The S1 replication's helps trigger is CI-clear by +0.0002 only, and its seed-0 cell reproduces the disclosed second smoke cell bit-exactly (a fresh training under the fresh-state pin; the F-18 accidental determinism check again — so the 0.0027 gap between the two smoke cells was the construction change, not CUDA jitter, a dated note on the block's SMOKE CONTACT sentence). Not re-examined at the corrected twin, which ran the gate cell only: the S1 offset-3 rows that carry OFF3-ABSORBED (the shift twin is CI-clear positive there, Mn off3 +0.0015 [+0.0005, +0.0025] / +0.0019 [+0.0000, +0.0038]), the S1 linear rows (replication Mt linear +0.0032 [+0.0010, +0.0055], upper above the floor, ABSORBED riding on D linear +0.0012 [-0.0002, +0.0026]) and the S3 primary linear row — those legs rest on the registered twin only. At linear capacity none of the blocks adds (S3 primary Mt linear +0.0003 [-0.0009, +0.0016], Mn linear -0.0012 [-0.0027, +0.0003]; S1 primary Mt linear -0.0002 [-0.0017, +0.0014]), so the +0.003 is mlp256-specific, consistent with a dims-or-optimization effect rather than telemetry content. The per-cluster correlations recorded in the twin payloads (aligned gain against twin gain, +0.58..+0.76) are a shared-subtrahend artifact — the three deltas share the fplus probe — and license nothing: cross-seed the correlations collapse to about +0.1 and a same-subtrahend control sits at +0.49 (critic-recomputed on the S3 primary; no committed artifact records it); the twin block's descriptive "shared row-level content" reading of them is withdrawn by dated note. The twin diagnostic's own pinned prediction FAILED in every part: branch (TWIN-ROBUST at both sources, predicted S1 TWIN-DEPENDENT), Mn_r sign (+0.0032 / +0.0033 CI-clear on the primaries, predicted near zero or slightly negative) and T (about zero everywhere, predicted +0.002..+0.003 on the S1 primary) — ROBUST arrived through a non-inert twin, not through the near-inert premise the block allowed for; the future-context mechanism the diagnostic was registered to price is withdrawn under its own pinned reading where the corrected twin helps CI-clear (both S1 harvests, the S3 primary), and is neither withdrawn nor supported on the S3 replication, where Mn, Mn_r and T all straddle zero. A weak free channel both twins share — the anchor's position — was measured by the critic on both primaries and is negligible (a position-only predictor gains nothing over the floor; no committed artifact records it). The S3 primary linear D row keeps its CI-positive-immaterial qualifier. (iv) THE READER: best VAL epochs on the h_a-bearing arms were 0-2 at every gate cell on every harvest (S1 primary 0 / 2 / 1 for all three arms; S1 replication 2 / 0 / 1; S3 primary fplus 1 1 1, fplus_b 1 0 1, fplus_n 1 1 1; S3 replication fplus 0 1 0, fplus_b 1 1 1, fplus_n 0 1 1; every S1 off3 cell 0) against 9-24 for the committed (a) mlp256 cells and 2-11 for (b)/(c) — the effective reader on these arms is one to three AdamW passes, a fact the registration disclosed from the smoke and both labels are scoped to. The dense standardized hz block dominates the input norm (per-anchor L2 about 65 against at most about 11 for the 128-hot slots block and about 6 for TOK on S1; critic-recomputed, no committed artifact records it), the one mechanism by which this reader could return a null a stronger reader would not — the stronger-reader residual, named under both labels as registered. (v) THE PREDICTION FAILED ON BOTH SOURCES: the pinned honest prediction was S1 SURVIVES-FULL-FREE with D +0.005..+0.012 and Mt positive but smaller than D, S3 SPLIT or NOT-SHOWN with ABSORBED "unlikely at this power". S1 read ABSORBED, the registered "what would surprise" branch; Mt exceeded D on every harvest; and S3's ABSORBED arrived because the D CIs are half the width of the Phase-3 (b)-(a) CIs (the twins share 82-86% of their dims plus the seeded init and batch order) and because the non-inert twin pushed D to zero by subtraction — not a power story. The registered parenthetical for the S1 surprise ("the S1 margin was nonlinear-free information all along") is quoted as pre-registered text, not asserted: the reader-handicap reading is equally consistent with the data. CONSEQUENCES EXECUTED (dated, per the pinned rules; never rewrites): F-15 and F-17 receive dated errata scoped to the re-tested cells (S1 L1 offsets 2-3; S3 L1 offset 2 — L2 / L3 and the further offsets stand against the 657-summary proxy and are untested against the full free baseline); K-2 receives a dated STATUS line per source; F-16 and F-18 receive dated "moot" notes. ABSORBED at BOTH sources, twin-robust at both: the existence layer is NOT SHOWN anywhere at the registered reader; K-2's 2026-09-05 STATUS salvage line ("the negative results and the method", docs/WHY_THIS_DIES.md K-2 STATUS 2026-09-05) is all that stands; scorecards get dated STATUS lines, never rewrites.
    sideband/FINDINGS.md:275 · d147e8c0cd6b
  17. 2026-10-16sidebandF-24dated note

    DATED NOTE 2026-10-16 (support-number artifacts; lever 2 of docs/PUBLIC_CUT_PLAN.md section 14.2, the rider of the S1 READER RE-TEST REGISTRATION in docs/BLUEPRINT.md; NON-CRITERION, no label moves; the three "no committed artifact records it" clauses above stay as written): …

    The full line, as written
    - DATED NOTE 2026-10-16 (support-number artifacts; lever 2 of docs/PUBLIC_CUT_PLAN.md section 14.2, the rider of the S1 READER RE-TEST REGISTRATION in docs/BLUEPRINT.md; NON-CRITERION, no label moves; the three "no committed artifact records it" clauses above stay as written): the three support figures are regenerated by committed scripts under docs/audits/2026-10-16-f24-support/ (each with its header pins set before the rerun and its output beside it; README.md there has the table): (1) corr_control.py reproduces the twin payloads' pooled per-cluster correlations to 0.0 on all four harvests and reads, on the S3 primary, the cross-seed Mt_i against Mn_r_j mean +0.1127 (the six ordered seed pairs -0.189..+0.511; within-seed mean +0.7837) and the same-subtrahend control +0.4815 under the construction the script names (Mt_i against acc(fplus_j) - acc(fplus_i), j != i); the record does not say which construction the critic used, and a second reading, pooled Mn against Mn_r, reads +0.4671 there, so "+0.49" is approximated by the first (+0.48 at two decimals) and within 0.03 by the second, and the first is the committed definition from this date; the recorded range "+0.58..+0.76" quoted in (iii) above covers the two S1 harvests and the S3 primary, the S3 replication's recorded values being +0.4062 / +0.4500 (the twin OUTCOME in docs/BLUEPRINT.md lists all four pairs; a dated correction, the sentence above stays as written); (2) position_only.py reads the per-index majority predictor at the gate cell equal to the constant majority predictor at every anchor index on both primaries (no index predicts a non-majority class; gain over the floor 0.0 exactly; floors 0.591381 / 0.580311 reproduced to 1e-12), the strongest form of "gains nothing" for one predictor (position jointly with the blocks is not what it measures); (3) input_norms.py reads, over the 61,030 S1 primary test anchors, the per-anchor L2 medians hz 65.21 (p10 57.27, p90 68.93, max 72.36), slots 11.14 (max 11.3137, sqrt(128) by construction), TOK 5.66 (max 5.6569, sqrt(32)) and (a) 2.23, with the hz moments sha equal to this entry's payload's; every pin held (numpy 2.3.5). [withheld: docs/PUBLIC_CUT_PLAN.md not in this repository; see WITHHELD.md]
    sideband/FINDINGS.md:277 · d147e8c0cd6b
  18. 2026-10-16sidebandF-28dated note

    DATED NOTE 2026-10-16 (image audit 2026-10-15, section B items 2 and 4; NON-CRITERION, no label moves at the registered estimator): …

    The full line, as written
    - DATED NOTE 2026-10-16 (image audit 2026-10-15, section B items 2 and 4; NON-CRITERION, no label moves at the registered estimator): (i) every interval of this entry is, like F-24's and F-25's, a prompt-cluster bootstrap over counts summed across three train seeds (per_seed_margin and n_seeds 3 in the payloads), so it is conditional on those three trained probes; the audit's seed-inclusive brackets (docs/audits/2026-10-15-image-audit/seed_var_out.json) cover F-24, its corrected twin and F-25 only, so no row of this entry is bracketed, the SURVIVES-STRONGER cells included. (ii) The nearest label-deciding bound to the +0.005 floor anywhere in the record is this entry's RN at L1 replication gate D +0.003206 [+0.001563, +0.004951] (results/real/readers_s3_r.json; primary +0.003035 [+0.001564, +0.004548]). The percentile cluster bootstrap covers about 0.939 at 30 clusters (docs/audits/2026-10-15-image-audit/coverage_out.json), and a t-for-z correction of the upper half-width (2.0452 / 1.96) carries that upper to +0.005026, at which the frozen rule (every D upper below +0.005 for ABSORBED) reads NOT-SHOWN for RN at L1 instead of ABSORBED (fed through the frozen tools/run_fullfree_coarse.py source_label: ABSORBED as registered, NOT-SHOWN with the scaled upper). The registered estimator is the percentile bootstrap, so the label stands as registered and is disclosed here as coverage-sensitive; the final S3 word K-2-ISOLATED does not change.
    sideband/FINDINGS.md:320 · d147e8c0cd6b
  19. 2026-10-11amberF-5erratum
    - erratum (2026-10-11): see F-16, row E17
    amber/FINDINGS.md:42 · 711fe31da2a4
  20. 2026-10-11amberF-7erratum
    - erratum (2026-10-11): see F-16, row E17
    amber/FINDINGS.md:58 · 711fe31da2a4
  21. 2026-10-11amberF-10erratum
    - erratum (2026-10-11): see F-16, row E19
    amber/FINDINGS.md:84 · 711fe31da2a4
  22. 2026-10-11amberF-11erratum
    - erratum (2026-10-11): see F-15, row E15
    amber/FINDINGS.md:93 · 711fe31da2a4
  23. 2026-10-11amberF-11erratum
    - erratum (2026-10-11): see F-16, row E18
    amber/FINDINGS.md:94 · 711fe31da2a4
  24. 2026-10-11amberF-12erratum
    - erratum (2026-10-11): see F-16, row E18
    amber/FINDINGS.md:104 · 711fe31da2a4
  25. 2026-10-11amberF-13erratum
    - erratum (2026-10-11): see F-16, row E18
    amber/FINDINGS.md:112 · 711fe31da2a4
  26. 2026-10-11amberF-14erratum
    - erratum (2026-10-11): see F-15, row E16
    amber/FINDINGS.md:138 · 711fe31da2a4
  27. 2026-10-11lockstepF-20dated note

    **Status**: acted on (TP-8). Harvested into sideband at e74cbd78e17d7a9065ea7cf7ab6e526174bc47b2 (branch `espalier-trial`): …

    The full line, as written
    **Status**: acted on (TP-8). Harvested into sideband at e74cbd78e17d7a9065ea7cf7ab6e526174bc47b2 (branch `espalier-trial`): `docs/DECODE_CONTRACTS.md` (the value contract's grid crossings beside amber's order-contract mismatches, every number generated from pinned commits), a dated addition to `docs/METHOD.md` section 6 (F-14's 114.5x pack-rate error), and dated notes in sideband's OQ-7, BLUEPRINT T-A, GLOSSARY and ECOSYSTEM saying that this repository's byproduct economics were never measured and its runtime never built. F-19's "no further findings will be appended" is amended by this entry, which is the last. Repository parked 2026-10-11.
    lockstep/trellis-docs/trellis-findings.md:188 · b9dff0132934
  28. 2026-10-10sidebandF-27amendment

    STAGE REPLAY (2026-10-10; OQ-20, the cross-seed replay of the two conversion-verdict legs and their mechanism-matched free legs at seed 6; NON-CRITERION for F-19, F-21 and this entry; registered 2026-10-09 in docs/BLUEPRINT.md, "SEED-REPLAY OF THE CONVERSION-VERDICT LEGS REGISTRATION", frozen at b6d3701): THE STAGE WORDS: …

    The full line, as written
    - STAGE REPLAY (2026-10-10; OQ-20, the cross-seed replay of the two conversion-verdict legs and their mechanism-matched free legs at seed 6; NON-CRITERION for F-19, F-21 and this entry; registered 2026-10-09 in docs/BLUEPRINT.md, "SEED-REPLAY OF THE CONVERSION-VERDICT LEGS REGISTRATION", frozen at b6d3701): THE STAGE WORDS: SPREAD-COVERED over R2 (K3 MUTUAL-COVERAGE / REPRODUCES, OQ15 MUTUAL-COVERAGE / REPRODUCES) and FREE-SPREAD-COVERED over R6 (K3 MUTUAL-COVERAGE / REPRODUCES, OQ15 MUTUAL-COVERAGE / REPRODUCES); said in words: each verdict's vs-vanilla leg and its mechanism-matched free leg, replayed at seed 6 under the frozen policies, reads inside the verdict seed's CI95 with the verdict's delta inside the seed-6 CI95, and each cross-seed difference's joint-draw CI95 contains zero (K3 -0.003479 [-0.033416, +0.022758] / -0.003135 [-0.033327, +0.023051]; OQ15 +0.020693 [-0.044830, +0.088277] / -0.028223 [-0.099002, +0.042346]); R2's and R6's CI95 half-widths on OQ15, 0.0666 and 0.0707, both exceed the +0.062 bar, so a cross-seed move the size of the bar is not resolved there, as the registration said it would not be (on K3 they are 0.0281 and 0.0282); the OQ15 vs-vanilla leg changes sign across the seeds (+0.002664 [-0.044719, +0.046727] at seed 3 against -0.018029 [-0.065168, +0.028495] at seed 6) with neither CI clear of zero, which is not the pinned surprise (that needs the seed-6 CI95 clear of zero); on K3 the free leg is nearly the vs-vanilla leg (|R5 - R1| = 0.000231), so the two stage words are not two confirmations there, as the registration said; the only leg whose seed-6 CI95 is clear of zero is OQ15's free leg, negative: the telemetry head sits below the free head at the new seed as at the verdict seed (D-5; no R1 point reads as telemetry help, no bar is cleared, K-3 stays FIRED); the guards held on every leg (margins +0.017778 / +0.060652 vs vanilla, +0.017509 / +0.041583 vs the free arm); no leg clears the bar at seed 6 (BAR-NOT-CLEARED and BAR-NOT-CLEARED-FREE on both); R4 R-CONSISTENT on both; R8 NEAR-IDENTICAL-ARMS on K3 (fixed before any run by the seed-1 count) and DELTA-MOVES / WITHIN-CROSS-SEED-NOISE on OQ15. ONE further seed per vs-vanilla leg and per mechanism-matched free leg; the cross-seed variance stays unestimated; the draft_conf legs and the same-seed null stay unmeasured; NON-CRITERION: no verdict, label or kill status moves. K3 (label, F-19; probe (c) telemetry scheduling (arm 4) against vanilla gamma-4, verdict seed 1): R1 the leg at seed 6 +0.010340 [-0.009288, +0.030521] GUARD-HELD (tokens/call 3.039368 vs 3.029028; drafted-per-emitted 1.312778 vs 1.320556, guard margin +0.017778); R2 MUTUAL-COVERAGE / REPRODUCES: the verdict leg +0.006860 [-0.015024, +0.025572] at seed 1 against +0.010340 [-0.009288, +0.030521] at seed 6, difference -0.003479 [-0.033416, +0.022758], in magnitude 0.17 of the verdict CI's half-width 0.0203 (the joint-draw CI95 half-width 0.0281; the seed-6 CI's 0.0199); R3 BAR-NOT-CLEARED; R4 R-CONSISTENT (the committed pair 0-1 r 0.4232 inside both new pairs' CI95: pair 0-6 r 0.4145 [0.2924, 0.5210], cross-seed null delta +0.002995 [-0.036426, +0.043831] (half-width 0.0401); pair 1-6 r 0.4036 [0.2857, 0.5172], cross-seed null delta -0.008601 [-0.048635, +0.031754] (half-width 0.0402); the pairs share runs, not independent replications); R5 the free leg at seed 6 (probe (c) telemetry scheduling (arm 4) minus probe (a) free scheduling (arm 3)) +0.010109 [-0.009724, +0.030422] GUARD-FREE-HELD (free arm tokens/call 3.029259, drafted-per-emitted 1.320287, guard margin +0.017509); R6 MUTUAL-COVERAGE / REPRODUCES: the verdict free leg +0.006974 [-0.014909, +0.025690] at seed 1 against +0.010109 [-0.009724, +0.030422] at seed 6, difference -0.003135 [-0.033327, +0.023051], in magnitude 0.15 of the verdict CI's half-width 0.0203 (the joint-draw CI95 half-width 0.0282); R7 BAR-NOT-CLEARED-FREE; R8 NEAR-IDENTICAL-ARMS: the free arm minus vanilla differs on 1 prompt at seed 1 and 3 at seed 6 in (n_emitted, n_blocks), below the pinned 10, so no label and no reading (only the counts are reported; the registration suppressed the label and reading for this case because the seed-1 leg's CI95 upper bound is exactly 0.0). OQ15 (replication, F-21; C4 head_tel (arm 7) against vanilla gamma-4, verdict seed 3): R1 the leg at seed 6 -0.018029 [-0.065168, +0.028495] GUARD-HELD (tokens/call 2.998372 vs 3.016401; drafted-per-emitted 1.275432 vs 1.326084, guard margin +0.060652); R2 MUTUAL-COVERAGE / REPRODUCES: the verdict leg +0.002664 [-0.044719, +0.046727] at seed 3 against -0.018029 [-0.065168, +0.028495] at seed 6, difference +0.020693 [-0.044830, +0.088277], in magnitude 0.45 of the verdict CI's half-width 0.0457 (the joint-draw CI95 half-width 0.0666; the seed-6 CI's 0.0468); R3 BAR-NOT-CLEARED; R4 R-CONSISTENT (the committed pair 2-3 r 0.4172 inside both new pairs' CI95: pair 2-6 r 0.4700 [0.3512, 0.5776], cross-seed null delta +0.009348 [-0.030631, +0.046660] (half-width 0.0386); pair 3-6 r 0.4736 [0.3691, 0.5698], cross-seed null delta -0.008126 [-0.044525, +0.030495] (half-width 0.0375); the pairs share runs, not independent replications); R5 the free leg at seed 6 (C4 head_tel (arm 7) minus C3 head_free (arm 6)) -0.053623 [-0.107526, -0.002448] GUARD-FREE-HELD (free arm tokens/call 3.051995, drafted-per-emitted 1.307015, guard margin +0.041583); R6 MUTUAL-COVERAGE / REPRODUCES: the verdict free leg -0.081846 [-0.132484, -0.035567] at seed 3 against -0.053623 [-0.107526, -0.002448] at seed 6, difference -0.028223 [-0.099002, +0.042346], in magnitude 0.58 of the verdict CI's half-width 0.0485 (the joint-draw CI95 half-width 0.0707); R7 BAR-NOT-CLEARED-FREE; R8 DELTA-MOVES / WITHIN-CROSS-SEED-NOISE: the free arm minus vanilla +0.084510 [+0.041028, +0.126063] at seed 3 against +0.035594 [-0.009909, +0.082906] at seed 6, difference +0.048916 [-0.011070, +0.110568] (the two arms differ on 210 / 210 prompts; a description, no bar, no guard, no surprise, no consequence beyond this line; docs/IDEAS.md rank 9 stays as open as it was). REGIME AND EVIDENCE: REAL MODELS — the Llama-3.1-8B-Instruct NF4 target with the Llama-3.2-1B-Instruct bf16 draft, temperature 1.0, 256 new tokens, the HF adapter (acceptance-valid only, D-9), one RTX 3080 Ti, seed 6 on the two 210-prompt verdict populations, the readouts a CPU recompute on the committed summaries; the six registered seed-6 runs of tools/run_k3_arms.py from a clean clone of the launch commit b6d3701 (every meta carries the bare commit, seed 6, one run_id, no resume; the pinned replay check PASSED on all six arms before any run, results/real/oq20_replay_check_stdout.txt its stdout and results/real/oq20_driver.log the launch driver's log; the environment capture after the last run equals the one before the first, results/real/oq20_env_capture_before_runs.txt and results/real/oq20_env_capture_after_runs.txt, sha256 f9bea69f380e... each), results/real/k3_seed6_arm1, k3_seed6_arm4, k3_seed6_arm3, oq15c_seed6_arm1, oq15c_seed6_arm7, oq15c_seed6_arm6 (five sidecars each, committed at c62f03d) beside the committed verdict runs and the F-27 payloads; the CPU readouts results/real/seed_replay_k3.json and seed_replay_oq15.json with their .state.jsonl, from the clean tree at c62f03d by tools/run_seed_replay.py (0b00f8dfd4cf as frozen; --conclude-from reproduces both payloads from the summaries on disk; the precondition held, 93 checks on K3 and 94 on OQ15); the critic's recompute (sideband-critic, opus, 2026-10-10: an independent parse of the per-prompt rows, the two sideband.stats estimators cross-checked against its own bootstrap under the same draws, every precondition re-verified from the metas, the policy files and the clone, both --conclude-from runs repeated in a fresh clone of the sidecars' commit) matched both payloads at a maximum absolute discrepancy of 0.0 over, by its own leaf count, 206 numeric leaves and 100 label and boolean leaves, confirmed K-3's R8 count (1 differing prompt at seed 1, 3 at seed 6) and found no leakage, pairing, regime or kill-criterion objection; its landing round was NO-GO on one blocking item outside the payloads, the frozen test file's pre-run assertions (the block's DATED AMENDMENT), with its advisories applied in these texts; its text round (same day) found one sign artifact and one false gloss in the generated sentences, both corrected by regeneration, and 0 misquoted numerals of 431; landing test tests/test_seed_replay_landing.py; tests/test_seed_replay.py amended at the landing by a dated line (48aec0c0590e; the block's DATED AMENDMENT); the block's OUTCOME scores both predictions. CONSEQUENCES EXECUTED (2026-10-10): this line; dated notes in docs/NOISE_FLOOR.md, README's "What is NOT measured" bullet, docs/REPORT.md's single-seed limitation and docs/FINDINGS_INDEX.md's F-27 row; NON-CRITERION notes in F-19's and F-21's status cells; the OUTCOME under OQ-20; the tool joins the frozen closure. Dated note 2026-10-16: the seed-6 arms replayed the same unclipped Stage-C heads; a counterfactual TRAIN-support clip changes 42 / 47 of their 17,157 / 17,602 consulted decisions (F-21's dated note of the same date; docs/audits/2026-10-15-image-audit/replay_c1_seed6.json). [withheld: docs/IDEAS.md not in this repository; see WITHHELD.md]
    sideband/FINDINGS.md:307 · d147e8c0cd6b
  29. 2026-10-10amberF-2erratum
    - erratum (2026-10-10): see F-14, row E12
    amber/FINDINGS.md:19 · 711fe31da2a4
  30. 2026-10-10amberF-4erratum
    - erratum (2026-10-10): see F-14, row E11
    amber/FINDINGS.md:34 · 711fe31da2a4
  31. 2026-10-10amberF-8erratum
    - erratum (2026-10-10): see F-14, rows E2, E3, E4, E5, E10
    amber/FINDINGS.md:66 · 711fe31da2a4
  32. 2026-10-10amberF-10erratum
    - erratum (2026-10-10): see F-14, rows E8a, E8c, E9
    amber/FINDINGS.md:82 · 711fe31da2a4
  33. 2026-10-10amberF-11erratum
    - erratum (2026-10-10): see F-14, rows E6, E7, E8b
    amber/FINDINGS.md:92 · 711fe31da2a4
  34. 2026-10-10amberF-14amendment

    E4 [CORRECTED; UNRECOMPUTABLE under the pinned rule, decided by the E4 amendment, 2026-10-10, written after the first dry run showed the float64 miss (instrument's float32 mean)] amber v0 0.4738 ties AC-under-draft 0.4723 bits/token (within 0.0015) (FINDINGS.md:59): …

    The full line, as written
    - E4 [CORRECTED; UNRECOMPUTABLE under the pinned rule, decided by the E4 amendment, 2026-10-10, written after the first dry run showed the float64 miss (instrument's float32 mean)] amber v0 0.4738 ties AC-under-draft 0.4723 bits/token (within 0.0015) (FINDINGS.md:59): the all-rows tie reproduces (amber 0.4738 vs AC-under-draft 0.4723 bits/token) and is a mixture: concatenated, non-looped rows (75 rows, 16958 tokens) amber 0.8929 vs AC 0.8094 (+0.0835); looped rows (125, 32000) 0.2515 vs 0.2937 (-0.0422). The per-archive direction depends on the loop rule: across the 39 grid settings the sign split holds at 39, while amber's per-archive rate on non-looped rows runs from -4.8% to +7.8% against AC (positive at 18), so no per-archive direction is stated.
    amber/FINDINGS.md:122 · 711fe31da2a4
  35. 2026-10-10amberF-14amendment

    E7 [CONFIRMED; UNRECOMPUTABLE under the pinned rule, decided by the E7 amendment, 2026-10-10, written after the first dry run could not load the tokenizer offline (loaded from the cached snapshot dir)] zstd-19 on the full text per archive: 7.28 bits per generated token (FINDINGS.md:80, OPEN_QUESTIONS.md:31): …

    The full line, as written
    - E7 [CONFIRMED; UNRECOMPUTABLE under the pinned rule, decided by the E7 amendment, 2026-10-10, written after the first dry run could not load the tokenizer offline (loaded from the cached snapshot dir)] zstd-19 on the full text per archive: 7.28 bits per generated token (FINDINGS.md:80, OPEN_QUESTIONS.md:31): reproduces (7.2805 bits per generated token with zstandard 0.25.0, tokenizer snapshot 4e20de36); the per-archive text includes decoded special-token literals (200 of 200 archives carry <\|begin_of_text\|>, 14 carry <\|end_of_text\|>, 3610 B in all); without them 6.75.
    amber/FINDINGS.md:125 · 711fe31da2a4
  36. 2026-10-10amberF-14amendment

    E13 [CORRECTED; UNRECOMPUTABLE under the pinned rule, decided by the E13 amendment, 2026-10-10, written after the fetched pages were read (citation_abstract instead of citation_title)] PMATIC is arXiv 2601.17684 (docs/PRIOR_ART.md:10): …

    The full line, as written
    - E13 [CORRECTED; UNRECOMPUTABLE under the pinned rule, decided by the E13 amendment, 2026-10-10, written after the fetched pages were read (citation_abstract instead of citation_title)] PMATIC is arXiv 2601.17684 (docs/PRIOR_ART.md:10): PMATIC is arXiv 2601.10678, "Synchronizing Probabilities in Model-Driven Lossless Compression" (Adler, Aviv; Tang, Jennifer; page sha256 a1ee3e4a2e0e); the cited id is arXiv 2601.17684, "A Model-Driven Lossless Compression Algorithm Resistant to Mismatch" (Hu, Cordelia; Tang, Jennifer; page sha256 b227437a106e), a different paper that is relevant prior art in its own right (re-identify the citation, do not remove it); pages saved under results/errata_web/.
    amber/FINDINGS.md:133 · 711fe31da2a4
  37. 2026-10-10lockstepF-19erratum
    **Context**: TP-7 (`task-packs/done/TP-7_errata_and_archive.md`), 2026-10-10; starting state 2d7d079; amber's errata pass at 0f6d2f0. Search command: `git grep -n -i lsp1 2d7d079 -- src`; output: 0 lines (empty).
    lockstep/trellis-docs/trellis-findings.md:178 · b9dff0132934
  38. 2026-10-04sidebandF-26amendment

    status (2026-10-04; HORIZON READ at the R2 encoding (K-6 lead-time pricing), the post-hoc diagnostic registered as the dated amendment "OQ-9 STAGE-1 HORIZON READ REGISTRATION" in docs/BLUEPRINT.md before any readout of it ran, frozen at b41d3f3; both runs from the clean tree at 5e0e2a9 (the next commit, a memory row); …

    The full line, as written
    - status (2026-10-04; HORIZON READ at the R2 encoding (K-6 lead-time pricing), the post-hoc diagnostic registered as the dated amendment "OQ-9 STAGE-1 HORIZON READ REGISTRATION" in docs/BLUEPRINT.md before any readout of it ran, frozen at b41d3f3; both runs from the clean tree at 5e0e2a9 (the next commit, a memory row); evidence results/real/prefetch_horizon_s3.json (primary) and results/real/prefetch_horizon_s3_r.json (replication), each with .meta.json and .state.jsonl; instrument tools/run_prefetch_horizon.py, which imports the frozen R2 tool and refuses a modified one; tests/test_prefetch_horizon.py, tests/test_prefetch_horizon_landing.py): LABEL LEAD-TIME-ABSORBED on the primary (H-ABSORBS / H-ABSORBS / H-ABSORBS at horizons 2, 3, 4; H-ABSORBS at horizon 1, R2's cell reproduced); the replication AGREES on the stage label and on every per-horizon label (a SEED replication on the same prompts; the primary's label is the label); no floor straddle, no reader guard and no free_h grid edge at any label horizon on either harvest, so the pinned cross-harvest closure condition holds. PROVENANCE: launched under the operator's words "all approved" (2026-10-04, to the proposed order: push the lane, the retroactive handoff, the primary then the replication, the landing lane); both runs from the clean tree at 5e0e2a9 (tracked status empty), the commit after the freeze commit b41d3f3 that added the session's memory row and nothing else (the eight files hashed at the freeze unchanged), sequential, instrument cc6a7fb0bc50..., the frozen chain at its shas, the replay source the pinned snapshot (sha d8e5770a2d88...), encoding prev_table with k and every R2 arm's lambda at horizon 1 taken from the committed R2 payloads; REPRODUCE EXACT against the R2 union rows (54,720 / 54,720 rows compared, 0 missing, 0 extra, 0 mismatches on hits or needed) and the identity on the four horizon-1 cells (D_eq^H, D_add^H, the anchor, free_h - free_r2 equal to R2's to 1e-12) on both; the horizon-1 RESELECT on val matches R2's on both (k = 128); 161,280 / 161,280 state rows; runtimes 2,940.3 / 1,842.0 s of elapsed time (per horizon 641.3 / 625.8 / 654.4 / 957.3 and 481.0 / 433.3 / 452.6 / 432.1; the replay 15.7 / 15.0 s of it); --conclude-from reproduces both payloads exactly; the test positions per horizon 59,285 / 59,045 / 58,805 / 58,565 and 29,825 / 29,705 / 29,585 / 29,465, the covered share 0.9013 / 0.9012 / 0.9011 / 0.9009 and 0.8582 / 0.8578 / 0.8573 / 0.8566 (the unseen-key share 0.0988 / 0.0989 / 0.0991 and 0.1422 / 0.1427 / 0.1434 at horizons 2, 3, 4). HORIZON 1 (H-ABSORBS; k = 128 on both, interior): D_eq^H (history_bigram_r2 - free_h, 3073 against 3073) +0.0111 [+0.0078, +0.0150] / +0.0106 [+0.0068, +0.0153] (NOT-SHOWN), D_add^H (history_all_h - free_h) +0.0172 [+0.0139, +0.0210] / +0.0162 [+0.0126, +0.0206] (NOT-SHOWN), the dilution anchor +0.0061 [+0.0057, +0.0064] / +0.0056 [+0.0052, +0.0060] (NOT-SHOWN), free_h - free_r2 +0.0438 [+0.0397, +0.0483] / +0.0450 [+0.0407, +0.0500] (NOT-SHOWN); ABT (history_tok - free_tok, the ADDRESS-BEYOND-TOKEN form at this horizon) +0.1033 [+0.0927, +0.1146] / +0.1074 [+0.0953, +0.1207] (SIGNAL); the twin's D_eq (history_tok - FREE_BIGRAM) +0.0609 [+0.0514, +0.0712] / +0.0659 [+0.0547, +0.0788] (SIGNAL) and D_add (HISTORY_BIGRAM - FREE_BIGRAM) +0.0661 [+0.0569, +0.0762] / +0.0705 [+0.0592, +0.0834] (SIGNAL); D_eq^R2 (history_tok_r2 - free_r2) +0.0493 [+0.0425, +0.0577] / +0.0506 [+0.0433, +0.0594] (NOT-SHOWN / SIGNAL) and D_add^R2 (history_bigram_r2 - free_r2) +0.0549 [+0.0480, +0.0631] / +0.0556 [+0.0483, +0.0645] (SIGNAL); h_only - free_tok -0.0051 [-0.0132, +0.0036] / +0.0035 [-0.0058, +0.0137] (NOT-SHOWN), h_only - history -0.0009 [-0.0082, +0.0057] / +0.0010 [-0.0072, +0.0078] (NOT-SHOWN), h_only - history_tok -0.1083 [-0.1129, -0.1039] / -0.1038 [-0.1090, -0.0992] (NEGATIVE); the label cells by scope: covered +0.0086 [+0.0059, +0.0121] / +0.0085 [+0.0055, +0.0125] (NOT-SHOWN) and +0.0147 [+0.0121, +0.0181] / +0.0141 [+0.0112, +0.0180] (NOT-SHOWN), fallback +0.0343 [+0.0213, +0.0465] / +0.0231 [+0.0131, +0.0342] (NOT-SHOWN) and +0.0399 [+0.0279, +0.0514] / +0.0286 [+0.0195, +0.0386] (NOT-SHOWN); ARMS at B = 16, pooled (primary / replication): marginal 0.4250 / 0.4312, history 0.8068 / 0.8060, free_tok 0.8110 / 0.8035, history_tok 0.9143 / 0.9109, FREE_BIGRAM 0.8534 / 0.8449, HISTORY_BIGRAM 0.9194 / 0.9155, free_r2 0.8732 / 0.8719, history_tok_r2 0.9225 / 0.9225, history_bigram_r2 0.9281 / 0.9275, h_only 0.8059 / 0.8071, free_h 0.9169 / 0.9169, history_all_h 0.9342 / 0.9331; lambda per arm (primary / replication where they differ): marginal 0.1, history 1000, free_tok 10, history_tok 10, FREE_BIGRAM 10, HISTORY_BIGRAM 10, free_r2 1, history_tok_r2 1, history_bigram_r2 1, h_only 100000, free_h 1, history_all_h 1; lambdas at a grid end on both harvests: free_replay (0.1), replay_cur (0.1), h_only (100000), marginal (0.1) (R2's taken lambdas carry no flag in this payload and are read against the grid's ends 0.1 and 100000; the marginal is val-selected here); B = 8 / 32 readings disagreeing with B = 16's on the label cells: none on either harvest; REPRODUCE EXACT, the identity held and the RESELECT matched on both (the horizon-1 arms are R2's readers and these cells are R2's cells). HORIZON 2 (H-ABSORBS; k = 128 on both, interior): D_eq^H (history_bigram_r2 - free_h, 3073 against 3073) +0.0034 [-0.0008, +0.0084] / +0.0024 [-0.0021, +0.0080] (NOT-SHOWN), D_add^H (history_all_h - free_h) +0.0178 [+0.0138, +0.0223] / +0.0153 [+0.0113, +0.0203] (NOT-SHOWN), the dilution anchor +0.0144 [+0.0135, +0.0152] / +0.0130 [+0.0120, +0.0140] (NOT-SHOWN), free_h - free_r2 +0.0677 [+0.0631, +0.0723] / +0.0683 [+0.0634, +0.0739] (SIGNAL); ABT (history_tok - free_tok, the ADDRESS-BEYOND-TOKEN form at this horizon) +0.1018 [+0.0912, +0.1132] / +0.1033 [+0.0910, +0.1171] (SIGNAL); the twin's D_eq (history_tok - FREE_BIGRAM) +0.0789 [+0.0693, +0.0889] / +0.0828 [+0.0713, +0.0955] (SIGNAL) and D_add (HISTORY_BIGRAM - FREE_BIGRAM) +0.0800 [+0.0705, +0.0899] / +0.0831 [+0.0717, +0.0958] (SIGNAL); D_eq^R2 (history_tok_r2 - free_r2) +0.0692 [+0.0614, +0.0784] / +0.0695 [+0.0613, +0.0797] (SIGNAL) and D_add^R2 (history_bigram_r2 - free_r2) +0.0711 [+0.0633, +0.0801] / +0.0707 [+0.0624, +0.0808] (SIGNAL); h_only - free_tok +0.0307 [+0.0226, +0.0392] / +0.0370 [+0.0288, +0.0461] (NOT-SHOWN), h_only - history +0.0110 [+0.0038, +0.0175] / +0.0130 [+0.0041, +0.0208] (NOT-SHOWN), h_only - history_tok -0.0711 [-0.0760, -0.0666] / -0.0663 [-0.0725, -0.0610] (NEGATIVE); the label cells by scope: covered +0.0002 [-0.0034, +0.0047] / -0.0001 [-0.0038, +0.0049] (NOT-SHOWN) and +0.0148 [+0.0115, +0.0188] / +0.0132 [+0.0099, +0.0178] (NOT-SHOWN), fallback +0.0333 [+0.0173, +0.0486] / +0.0172 [+0.0067, +0.0272] (NOT-SHOWN) and +0.0455 [+0.0318, +0.0582] / +0.0282 [+0.0197, +0.0368] (NOT-SHOWN); ARMS at B = 16, pooled (primary / replication): marginal 0.4254 / 0.4315, history 0.6680 / 0.6664, free_tok 0.6482 / 0.6424, history_tok 0.7500 / 0.7456, FREE_BIGRAM 0.6712 / 0.6629, HISTORY_BIGRAM 0.7512 / 0.7460, free_r2 0.6862 / 0.6830, history_tok_r2 0.7554 / 0.7525, history_bigram_r2 0.7573 / 0.7538, h_only 0.6790 / 0.6794, free_h 0.7539 / 0.7514, history_all_h 0.7717 / 0.7667; lambda per arm (primary / replication where they differ): marginal 0.1, history 1000, free_tok 10, history_tok 10, FREE_BIGRAM 10, HISTORY_BIGRAM 10, free_r2 1, history_tok_r2 1, history_bigram_r2 1, h_only 100000 / 10000, free_h 1, history_all_h 1; lambdas at a grid end: marginal (0.1), h_only (100000) on the primary, marginal (0.1) on the replication; B = 8 / 32 readings disagreeing with B = 16's on the label cells: none on either harvest. HORIZON 3 (H-ABSORBS; k = 128 on both, interior): D_eq^H (history_bigram_r2 - free_h, 3073 against 3073) +0.0016 [-0.0032, +0.0070] / +0.0002 [-0.0052, +0.0063] (NOT-SHOWN), D_add^H (history_all_h - free_h) +0.0173 [+0.0129, +0.0223] / +0.0135 [+0.0093, +0.0187] (NOT-SHOWN), the dilution anchor +0.0158 [+0.0148, +0.0166] / +0.0134 [+0.0117, +0.0149] (NOT-SHOWN), free_h - free_r2 +0.0740 [+0.0682, +0.0800] / +0.0736 [+0.0667, +0.0805] (SIGNAL); ABT (history_tok - free_tok, the ADDRESS-BEYOND-TOKEN form at this horizon) +0.1006 [+0.0890, +0.1127] / +0.0996 [+0.0863, +0.1141] (SIGNAL); the twin's D_eq (history_tok - FREE_BIGRAM) +0.0831 [+0.0728, +0.0940] / +0.0849 [+0.0726, +0.0983] (SIGNAL) and D_add (HISTORY_BIGRAM - FREE_BIGRAM) +0.0830 [+0.0728, +0.0937] / +0.0842 [+0.0721, +0.0975] (SIGNAL); D_eq^R2 (history_tok_r2 - free_r2) +0.0743 [+0.0655, +0.0839] / +0.0734 [+0.0641, +0.0845] (SIGNAL) and D_add^R2 (history_bigram_r2 - free_r2) +0.0755 [+0.0668, +0.0852] / +0.0737 [+0.0644, +0.0847] (SIGNAL); h_only - free_tok +0.0743 [+0.0659, +0.0831] / +0.0774 [+0.0691, +0.0868] (SIGNAL), h_only - history +0.0064 [-0.0003, +0.0122] / +0.0072 [-0.0015, +0.0142] (NOT-SHOWN), h_only - history_tok -0.0263 [-0.0322, -0.0212] / -0.0221 [-0.0294, -0.0159] (NEGATIVE); the label cells by scope: covered -0.0030 [-0.0067, +0.0016] / -0.0037 [-0.0075, +0.0011] (NOT-SHOWN) and +0.0138 [+0.0104, +0.0180] / +0.0106 [+0.0075, +0.0148] (NOT-SHOWN), fallback +0.0433 [+0.0188, +0.0683] / +0.0231 [+0.0057, +0.0412] (NOT-SHOWN) and +0.0499 [+0.0290, +0.0710] / +0.0309 [+0.0175, +0.0449] (NOT-SHOWN); ARMS at B = 16, pooled (primary / replication): marginal 0.4259 / 0.4320, history 0.5951 / 0.5944, free_tok 0.5272 / 0.5241, history_tok 0.6278 / 0.6237, FREE_BIGRAM 0.5448 / 0.5388, HISTORY_BIGRAM 0.6278 / 0.6230, free_r2 0.5584 / 0.5561, history_tok_r2 0.6327 / 0.6295, history_bigram_r2 0.6340 / 0.6298, h_only 0.6015 / 0.6015, free_h 0.6324 / 0.6296, history_all_h 0.6497 / 0.6431; lambda per arm (primary / replication where they differ): marginal 0.1, history 1000, free_tok 10, history_tok 10, FREE_BIGRAM 10, HISTORY_BIGRAM 10, free_r2 1, history_tok_r2 1, history_bigram_r2 10, h_only 100000 / 10000, free_h 10, history_all_h 10; lambdas at a grid end: marginal (0.1), h_only (100000) on the primary, marginal (0.1) on the replication; B = 8 / 32 readings disagreeing with B = 16's on the label cells: none on either harvest. HORIZON 4 (H-ABSORBS; k = 128 on both, interior): D_eq^H (history_bigram_r2 - free_h, 3073 against 3073) -0.0002 [-0.0052, +0.0056] / -0.0023 [-0.0075, +0.0039] (NOT-SHOWN), D_add^H (history_all_h - free_h) +0.0149 [+0.0105, +0.0198] / +0.0104 [+0.0062, +0.0154] (NOT-SHOWN), the dilution anchor +0.0151 [+0.0138, +0.0164] / +0.0128 [+0.0111, +0.0142] (NOT-SHOWN), free_h - free_r2 +0.0752 [+0.0681, +0.0824] / +0.0741 [+0.0665, +0.0824] (SIGNAL); ABT (history_tok - free_tok, the ADDRESS-BEYOND-TOKEN form at this horizon) +0.0942 [+0.0823, +0.1069] / +0.0923 [+0.0784, +0.1074] (SIGNAL); the twin's D_eq (history_tok - FREE_BIGRAM) +0.0808 [+0.0701, +0.0920] / +0.0814 [+0.0684, +0.0954] (SIGNAL) and D_add (HISTORY_BIGRAM - FREE_BIGRAM) +0.0806 [+0.0698, +0.0917] / +0.0809 [+0.0681, +0.0947] (SIGNAL); D_eq^R2 (history_tok_r2 - free_r2) +0.0742 [+0.0639, +0.0852] / +0.0713 [+0.0604, +0.0837] (SIGNAL) and D_add^R2 (history_bigram_r2 - free_r2) +0.0750 [+0.0647, +0.0860] / +0.0718 [+0.0609, +0.0841] (SIGNAL); h_only - free_tok +0.0862 [+0.0768, +0.0958] / +0.0879 [+0.0790, +0.0980] (SIGNAL), h_only - history +0.0032 [-0.0033, +0.0088] / +0.0034 [-0.0047, +0.0100] (NOT-SHOWN), h_only - history_tok -0.0080 [-0.0141, -0.0027] / -0.0043 [-0.0124, +0.0023] (NEGATIVE / NOT-SHOWN); the label cells by scope: covered -0.0050 [-0.0087, -0.0002] / -0.0060 [-0.0098, -0.0011] (NEGATIVE) and +0.0112 [+0.0079, +0.0155] / +0.0079 [+0.0047, +0.0122] (NOT-SHOWN), fallback +0.0431 [+0.0164, +0.0689] / +0.0196 [+0.0042, +0.0373] (NOT-SHOWN) and +0.0483 [+0.0264, +0.0693] / +0.0254 [+0.0134, +0.0390] (NOT-SHOWN); ARMS at B = 16, pooled (primary / replication): marginal 0.4260 / 0.4321, history 0.5586 / 0.5583, free_tok 0.4757 / 0.4738, history_tok 0.5698 / 0.5661, FREE_BIGRAM 0.4890 / 0.4847, HISTORY_BIGRAM 0.5696 / 0.5656, free_r2 0.5004 / 0.4995, history_tok_r2 0.5746 / 0.5708, history_bigram_r2 0.5755 / 0.5713, h_only 0.5618 / 0.5617, free_h 0.5757 / 0.5736, history_all_h 0.5906 / 0.5840; lambda per arm (primary / replication where they differ): marginal 0.1, history 1000, free_tok 10, history_tok 10, FREE_BIGRAM 10, HISTORY_BIGRAM 100, free_r2 10, history_tok_r2 10, history_bigram_r2 10, h_only 100000 / 10000, free_h 10, history_all_h 10; lambdas at a grid end: marginal (0.1), h_only (100000) on the primary, marginal (0.1) on the replication; B = 8 / 32 readings disagreeing with B = 16's on the label cells: none on either harvest. TREND (nested, different position sets; a description, no reading): D_add^H at horizons 1..4 +0.0172 / +0.0178 / +0.0173 / +0.0149 on the primary and +0.0162 / +0.0153 / +0.0135 / +0.0104 on the replication (the points at horizon 4 below those at horizon 2 on both, the assistant's surprise item; a description on nested position sets, no reading); D_eq^H +0.0111 / +0.0034 / +0.0016 / -0.0002 and +0.0106 / +0.0024 / +0.0002 / -0.0023, with zero inside the CI95 at horizons 2, 3 and 4 on both harvests (at horizon 1, R2's cell, it is CI-positive); free_h - free_r2 +0.0438 / +0.0677 / +0.0740 / +0.0752 and +0.0450 / +0.0683 / +0.0736 / +0.0741, larger at each later horizon on both; h_only - free_tok -0.0051 / +0.0307 / +0.0743 / +0.0862 and +0.0035 / +0.0370 / +0.0774 / +0.0879 (CI-positive at horizons 2, 3 and 4 on both). The horizon-matched ABT against stage 1's committed cell at the same horizon (a description; stage 1's positions include t = 0 and its one-token table is not horizon-matched): horizon 2 +0.1018 / +0.1033 against stage 1's +0.1027 / +0.1035 (below / below); horizon 3 +0.1006 / +0.0996 against stage 1's +0.1005 / +0.0995 (above / above); horizon 4 +0.0942 / +0.0923 against stage 1's +0.0933 / +0.0912 (above / above); the horizon-matched free_tok against stage 1's one-token table: horizon 2 0.6482 / 0.6424 against 0.6449 / 0.6402 (above / above); horizon 3 0.5272 / 0.5241 against 0.5303 / 0.5276 (below / below); horizon 4 0.4757 / 0.4738 against 0.4804 / 0.4796 (below / below); history against stage 1's: horizon 2 0.6680 / 0.6664 against 0.6677 / 0.6661; horizon 3 0.5951 / 0.5944 against 0.5950 / 0.5943; horizon 4 0.5586 / 0.5583 against 0.5586 / 0.5584. SAID IN WORDS, beside the labels: no floor straddle, no reader guard and no free_h grid edge at any horizon on either harvest, and no NEGATIVE reading on a label cell on ALL (the label scope); NEGATIVE readings, every one said in words: D_eq^H on COVERED at horizon 4 on both harvests; h_only - free_tok on COVERED at horizon 1 on both harvests; h_only - history on FALLBACK at horizons 1, 3, 4 on both harvests and horizon 2 on the primary only; h_only - history_tok on ALL at horizons 1, 2, 3 on both harvests and horizon 4 on the primary only; h_only - history_tok on COVERED at horizons 1, 2, 3 on both harvests; h_only - history_tok on FALLBACK at horizons 1, 2, 3, 4 on both harvests; no B = 8 / 32 reading disagrees with B = 16's on a label cell; h_only is not above history_tok at any horizon; no FALLBACK-scope label cell reads SIGNAL; the lambda grid edges fall on REPORTED arms only: the marginal at the grid's lower edge at every horizon on both harvests (a bias-only arm); h_only at the upper edge on the primary at every horizon (its lambda 100000 / 10000, primary / replication, at horizons 2..4: the grid's upper edge on the primary, interior on the replication; at horizon 1 both carry R2's 100000, the upper edge); and at horizon 1, on both harvests, R2's taken lambdas at a grid end, free_replay (0.1) and replay_cur (0.1), h_only (100000), and the val-selected marginal (0.1) at the lower edge (said as the registration requires: a grid-edge lambda on any arm at any horizon); the horizon-matched ABT sits below stage 1's committed cell at horizon 2 and above it at horizons 3 and 4 on both harvests, by at most 0.0011 (a description); the primary's unseen-key CI95 contains the floor on 5 of the six fallback label cells, the replication's on none (the fallback cells are reported, not label cells: the label reads ALL); the RESELECT at horizon 1 matches R2's on both. THE PINNED SENTENCE, in its words with the numbers beside the slots (primary / replication): "with the forward's final-norm state at the address block's width and two tokens' smoothed routing profiles in the free arm, no material margin (at or above +0.05) of the address block over the equal-dims free block (horizon 2 +0.0034 [-0.0008, +0.0084] / +0.0024 [-0.0021, +0.0080]; horizon 3 +0.0016 [-0.0032, +0.0070] / +0.0002 [-0.0052, +0.0063]; horizon 4 -0.0002 [-0.0052, +0.0056] / -0.0023 [-0.0075, +0.0039]) and none of the address stream over the free arm (horizon 2 +0.0178 [+0.0138, +0.0223] / +0.0153 [+0.0113, +0.0203]; horizon 3 +0.0173 [+0.0129, +0.0223] / +0.0135 [+0.0093, +0.0187]; horizon 4 +0.0149 [+0.0105, +0.0198] / +0.0104 [+0.0062, +0.0154]) is shown at horizon 2, 3 or 4, as at horizon 1 (R2): at this reader the address stream is not shown to add a material hit-rate margin over free information at any horizon 1..4 (lead time counted in forwards; no latency priced; NON-CRITERION; K-2 is untouched; K-6 stays UNTESTED and this is its lead-time datum); the stage-2 strike stands, PREFETCH-HEADROOM (which does not depend on this margin) stands, and OQ-9's prefetch half closes at this reader" (every one of the six label readings NOT-SHOWN by the rule; CI-positive below the floor for D_add^H at every horizon on both, for D_eq^H at horizon 1 only) — with the per-horizon FALLBACK readings of the two label cells beside it, as pinned (horizon 2: D_eq^H +0.0333 [+0.0173, +0.0486] / +0.0172 [+0.0067, +0.0272] (NOT-SHOWN), D_add^H +0.0455 [+0.0318, +0.0582] / +0.0282 [+0.0197, +0.0368] (NOT-SHOWN; the primary's CI95 containing the floor); horizon 3: D_eq^H +0.0433 [+0.0188, +0.0683] / +0.0231 [+0.0057, +0.0412] (NOT-SHOWN; the primary's CI95 containing the floor), D_add^H +0.0499 [+0.0290, +0.0710] / +0.0309 [+0.0175, +0.0449] (NOT-SHOWN; the primary's CI95 containing the floor); horizon 4: D_eq^H +0.0431 [+0.0164, +0.0689] / +0.0196 [+0.0042, +0.0373] (NOT-SHOWN; the primary's CI95 containing the floor), D_add^H +0.0483 [+0.0264, +0.0693] / +0.0254 [+0.0134, +0.0390] (NOT-SHOWN; the primary's CI95 containing the floor); the unseen-key share 0.0988 / 0.0989 / 0.0991 and 0.1422 / 0.1427 / 0.1434), and the COVERED readings (horizon 2: D_eq^H +0.0002 [-0.0034, +0.0047] / -0.0001 [-0.0038, +0.0049] (NOT-SHOWN), D_add^H +0.0148 [+0.0115, +0.0188] / +0.0132 [+0.0099, +0.0178] (NOT-SHOWN); horizon 3: D_eq^H -0.0030 [-0.0067, +0.0016] / -0.0037 [-0.0075, +0.0011] (NOT-SHOWN), D_add^H +0.0138 [+0.0104, +0.0180] / +0.0106 [+0.0075, +0.0148] (NOT-SHOWN); horizon 4: D_eq^H -0.0050 [-0.0087, -0.0002] / -0.0060 [-0.0098, -0.0011] (NEGATIVE), D_add^H +0.0112 [+0.0079, +0.0155] / +0.0079 [+0.0047, +0.0122] (NOT-SHOWN); NEGATIVE readings among the label cells, said in words: D_eq^H on COVERED at horizon 4 on both harvests). NON-CRITERION: no label of F-26, the twin, the stratified read or R2 moves; this read is NOT a rung on the horizon-1 margin (R2 closed that ladder) and no free-baseline rung is registered at any horizon; the address-beyond-free motivation struck at horizon 1 stays struck at horizons 2, 3 and 4; F-26's post-hoc reading (2) (the address margin over the one-token free arm holding across horizons 1..4) is reproduced here as a description (the horizon-matched ABT above) and does not carry over to the free arm R2 built: against free_h no material margin of the address stream is shown at any horizon, while free_h - free_r2 grows with the horizon (numbers above; a description, no mechanism claimed); no wall-clock, bytes or latency sentence (D-9; K-5 / K-6 stay UNTESTED: lead time is counted in forwards). PREDICTIONS SCORED by the registration's SCORING rule under five conventions, named here and applied to both lists alike: (i) every range and every modal label is scored HELD / FAILED, a modal label on its label word alone; (ii) a clause that carries an explicit probability ('about p', '(p)') is calibration and is not scored, whatever it predicts (among them the assistant's (9) and the critic's 1, the 'no reader guard (about 0.93)' half of its 9 and its 13; what they predicted is stated beside the labels above); (i) applies before (ii): a modal label is one the clause calls modal or the most likely label of a stated distribution over the stage label (the assistant's (6), the critic's 12), while a reading word such as SIGNAL with a probability beside it is not a modal label and falls under (ii) (the critic's 6); (iii) a clause is scored on the primary harvest, the label harvest, unless it names the replication; (iv) a directional header claim ('falls', 'rises') is scored as a bool on the points; (v) 'about half' is read as 0.40..0.60 and 'within' as a closed interval. THE ASSISTANT'S, 23 of 34 clauses HELD: (1) REPRODUCE EXACT on both, the identity holds, the horizon-1 RESELECT matches R2's on both -> HELD; (2) every arm's hit rate falls with the horizon (the eleven arms beside the marginal, horizons 1..4, strictly) -> HELD; (2) history_all_h at j = 2 in 0.80..0.86 -> +0.7717 FAILED; (2) history_all_h at j = 3 in 0.70..0.78 -> +0.6497 FAILED; (2) history_all_h at j = 4 in 0.64..0.72 -> +0.5906 FAILED; (2) the marginal in 0.42..0.44 at every horizon -> HELD; (2) free_h within 0.03 of history_all_h at j = 2 -> +0.0178 HELD; (2) free_h within 0.05 of history_all_h at j = 4 -> +0.0149 HELD; (2) h_only above free_tok at every j >= 2 -> HELD; (2) h_only - free_tok at j = 4 in +0.02..+0.10 -> +0.0862 HELD; (3) D_eq^H at j = 2 in +0.00..+0.03 -> +0.0034 HELD; (3) D_eq^H at j = 3 in +0.00..+0.04 -> +0.0016 HELD; (3) D_eq^H at j = 4 in +0.00..+0.05 -> -0.0002 FAILED; (3) D_eq^H NOT-SHOWN modal at every horizon -> HELD; (4) D_add^H rises with j -> FAILED; (4) D_add^H at j = 2 in +0.015..+0.035 -> +0.0178 HELD; (4) D_add^H at j = 3 in +0.02..+0.05 -> +0.0173 FAILED; (4) D_add^H at j = 4 in +0.02..+0.06 -> +0.0149 FAILED; (4) D_add^H NOT-SHOWN modal at j = 2 and 3 -> HELD; (5) the anchor in +0.00..+0.02 and CI-positive at every horizon -> HELD; (5) free_h - free_r2 CI-positive and growing with j -> HELD; (5) free_h - free_r2 at j = 4 in +0.06..+0.15 -> +0.0752 HELD; (5) no reader guard at any horizon -> HELD; (6) STAGE LEAD-TIME-ABSORBED modal -> HELD; (7) ABT at j = 2 below stage 1's cell by 0.01..0.04 -> +0.0009 FAILED; (7) ABT at j = 3 below stage 1's cell by 0.01..0.04 -> -0.0001 FAILED; (7) ABT at j = 4 below stage 1's cell by 0.01..0.04 -> -0.0009 FAILED; (7) ABT SIGNAL at every horizon -> HELD; (8) k at j >= 2 interior and no larger than 128 -> HELD; (8) free_r2's lambda interior at every horizon -> HELD; (8) h_only's lambda at the grid's upper edge (the primary; the replication's 100000 / 10000 at horizons 2..4 is interior, said in words) -> HELD; (10) 30..50 minutes on the primary (49.0 min) -> HELD; (10) about half on the replication (scored as 0.40..0.60 of the primary; 0.6265) -> FAILED; (10) 150,000 to 175,000 state rows per harvest (161,280 / 161,280) -> HELD. Surprise list: 2 fired (D_add^H falling from j = 2 to j = 4 (the primary); the horizon-matched ABT above stage 1's at some j >= 2 (the primary, horizons 3, 4, by +0.0001 / +0.0009)). THE CRITIC'S, 23 of 34 clauses HELD: 2. history at j = 2, 3, 4 within 0.004 of stage 1's committed cells -> HELD; 3. the marginal in 0.420..0.432 on the primary at every horizon -> HELD; 3. the marginal in 0.426..0.437 on the replication at every horizon -> HELD; 4. horizon-matched free_tok above stage 1's at every j >= 2 -> FAILED; 4. free_tok at j = 2 in 0.66..0.71 -> +0.6482 FAILED; 4. free_tok at j = 3 in 0.55..0.61 -> +0.5272 FAILED; 4. free_tok at j = 4 in 0.50..0.56 -> +0.4757 FAILED; 5. history_all_h at j = 2 in 0.765..0.82 -> +0.7717 HELD; 5. history_all_h at j = 3 in 0.65..0.715 -> +0.6497 FAILED; 5. history_all_h at j = 4 in 0.59..0.66 -> +0.5906 HELD; 6. ABT at j = 2 in +0.06..+0.095 -> +0.1018 FAILED; 6. ABT at j = 3 in +0.055..+0.095 -> +0.1006 FAILED; 6. ABT at j = 4 in +0.045..+0.09 -> +0.0942 FAILED; 6. ABT below stage 1's cell at every j >= 2 -> FAILED; 7. D_eq^H at j = 2 in +0.000..+0.030 -> +0.0034 HELD; 7. D_eq^H at j = 3 in +0.000..+0.035 -> +0.0016 HELD; 7. D_eq^H at j = 4 in -0.005..+0.035 -> -0.0002 HELD; 7. D_eq^H NOT-SHOWN modal at every horizon -> HELD; 8. D_add^H at j = 2 in +0.012..+0.035 -> +0.0178 HELD; 8. D_add^H at j = 3 in +0.012..+0.038 -> +0.0173 HELD; 8. D_add^H at j = 4 in +0.010..+0.040 -> +0.0149 HELD; 8. D_add^H NOT-SHOWN modal -> HELD; 8. D_add^H(4) - D_add^H(2) within 0.012 -> -0.0029 HELD; 9. the anchor in +0.003..+0.015 and CI-positive at every j >= 2 -> FAILED; 9. free_h - free_r2 in +0.025..+0.07 and CI-positive at every j >= 2 -> FAILED; 10. h_only - history at j = 4 in -0.04..+0.01 -> +0.0032 HELD; 11. k at j >= 2 in {128, 512} modal -> HELD; 11. free_r2's lambda interior everywhere -> HELD; 11. h_only's lambda 1e4 or 1e5 at every j >= 2 -> HELD; 12. STAGE LEAD-TIME-ABSORBED modal -> HELD; 14. exactly 161,280 state rows per harvest -> HELD; 14. the covered share at j >= 2 within -0.012..0.000 of j = 1's on both -> HELD; 15. 35..60 minutes on the primary (49.0 min) -> HELD; 15. the replication 0.45..0.75 of the primary (0.6265) -> HELD. Surprise list: 1 fired (horizon-matched free_tok below stage 1's at some j >= 2 (the primary, horizons 3 and 4)). The critic's 'D_eq^H NEGATIVE at any horizon' is read on ALL, the label scope, where it does not fire; on the COVERED scope D_eq^H reads NEGATIVE at horizon 4 on both harvests (said above). Critic recompute (2026-10-04, recompute mode, an independent numpy script over the two state files and the committed R2 state files, dispatched at 17:48 local and read before the first edit): GO at maximum absolute discrepancy 0.0 over 75,816 payload values (both harvests), 1,188 identity values and 218,880 REPRODUCE integer comparisons, the selections and every label and flag re-derived with 0 mismatches; the record in the block's OUTCOME. CLEAN-CLONE RERUN (the landing condition under CONSEQUENCES (iv); started at 17:47 local in a scratch clone, in parallel with the recompute critic, whose GO came at 18:04; its failure branch pinned at 18:11, before its state file was read at 18:38): the primary rerun from a fresh clone of the freeze commit b41d3f3 (the clone's own tool, R2 payload and state file, all -text; the gitignored harvest shards copied from the main tree and checked by the tool against the committed metas and the R2 record; the replay from the cache) to the clone's own results path, 3,017.5 s of elapsed time on the same box and BLAS (scipy-openblas 0.3.30; the landing tests ran beside it): BYTE-IDENTICAL state file (sha256 e768ca7a68db..., the registered primary's), and the conclusion, the k and lambda val tables, the selections, the hidden block's spectrum, the replay probe, the strata counts and the coverage records all equal; the provenance differs in the commit it attests (b41d3f3, the freeze itself), the paths and the timings only — so the val tables that decide k and every lambda, which no committed row recomputes, reproduce on this box from the freeze commit, and the pinned failure branch was not needed. Both predictions scored in the block's OUTCOME, with the recompute record.
    sideband/FINDINGS.md:298 · d147e8c0cd6b
  39. 2026-10-03sidebandF-26amendment

    status (2026-10-03; GENERALISING FREE BASELINE (R2), the post-hoc diagnostic registered as the dated amendment "OQ-9 STAGE-1 GENERALISING FREE BASELINE REGISTRATION (R2)" in docs/BLUEPRINT.md before any readout of it ran, both runs from the clean freeze commit 2fe5926; …

    The full line, as written
    - status (2026-10-03; GENERALISING FREE BASELINE (R2), the post-hoc diagnostic registered as the dated amendment "OQ-9 STAGE-1 GENERALISING FREE BASELINE REGISTRATION (R2)" in docs/BLUEPRINT.md before any readout of it ran, both runs from the clean freeze commit 2fe5926; evidence results/real/prefetch_r2_s3.json (primary) and results/real/prefetch_r2_s3_r.json (replication), each with .meta.json and .state.jsonl; instrument tools/run_prefetch_r2.py, which imports the frozen stratified tool and refuses a modified one; tests/test_prefetch_r2.py, tests/test_prefetch_r2_landing.py): LABEL R2-FALLBACK-SURVIVES / H-ABSORBS on the primary; the replication AGREES on both questions (a SEED replication on the same prompts; the primary's label is the label). PROVENANCE: both runs from the clean freeze commit 2fe5926 (tracked status empty), sequential, instrument eaf671fdb9d6..., the frozen chain at its shas, the replay source the pinned snapshot (sha d8e5770a2d88...), encoding prev_table and the twin's lambdas from the committed stratified payloads; REPRODUCE EXACT against the stratified state rows (76,464 / 77,328 rows, 0 mismatches) and the identity on the three D_eq cells held; 161,424 / 163,248 state rows; runtimes 648.5 / 323.3 s of elapsed time (the replay 11.4 / 11.2 s of it); --conclude-from reproduces both payloads exactly; the test positions 59,285 / 29,825, the unseen-key share 0.0987 / 0.1418, the covered share 0.9013 / 0.8582. THE FREE ENCODING: the twin's count tables with the model's own routers replayed on the token's normalised input embedding as the Dirichlet prior of each key's row, one weight k selected on val jointly with lambda from {0, 2, 8, 32, 128, 512, inf}; SELECTION: k = 128 on both harvests (interior; the k val table's best rises from k = 0 to k = 128 and falls at 512 and inf; in occurrence units k = 128 is 16 virtual train occurrences per gate, so a key seen five times in train still reads 0.7619 replay; the selection margin over k = 32 is 1.3e-04 / 8.4e-04 of pooled val hit rate — the thinnest of the k and lambda selections in the primary run, re-derived by --conclude-from from the val tables in the meta but recomputable from no committed row; on the replication these lambda selections are thinner than the replication's own k margin (8.4e-04), by the same runner-up gap: h_only 1.3e-04, history_tok_r2 4.1e-04, history_all_h 7.5e-04, history_bigram_r2 8.2e-04), free_r2's lambda 1 / 1 (interior); GRID EDGES on three REPORTED arms, said in words: free_replay and replay_cur at lambda 0.1 / 0.1 (the grid's lower edge) and h_only at lambda 100000 / 100000 (the upper edge: the hidden block alone wants more regularisation than the grid offers, which bounds h_only's point, not free_h's — free_h's lambda 1 / 1 and history_tok_r2's 1 / 1 are interior). B = 8 / 32 readings that disagree with B = 16's on the four label cells, said in words: primary gate_eq_r2|B=32|fallback: NOT-SHOWN; gate_eq_r2|B=8|all: SIGNAL (+0.0262 [+0.0203, +0.0325]; +0.0671 [+0.0609, +0.0739]); replication gate_eq_r2|B=32|fallback: NOT-SHOWN; gate_eq_r2|B=32|all: NOT-SHOWN (+0.0232 [+0.0169, +0.0303]; +0.0156 [+0.0125, +0.0196]); no limit arm above free_r2 on test pooled; replay_zerofit 0.5675 / 0.5707 above stage 1's RECENCY_16 0.5197 / 0.5155 (the comparator does not fire; the position sets differ, stage 1's j = 1 positions including t = 0). SHRINKAGE (free_r2 - FREE_BIGRAM) on the unseen-key positions +0.1971 [+0.1697, +0.2266] / +0.1822 [+0.1557, +0.2167] (SIGNAL; per stratum cur-unseen +0.2604 [+0.2433, +0.2771] / +0.2608 [+0.2425, +0.2778] (SIGNAL), prev-unseen +0.0585 [+0.0488, +0.0691] / +0.0473 [+0.0401, +0.0544] (SIGNAL / NOT-SHOWN), both-unseen +0.4895 [+0.4194, +0.5142] / +0.4402 [+0.3582, +0.4684] (SIGNAL)), on the covered positions +0.0004 [-0.0005, +0.0013] / +0.0013 [+0.0004, +0.0023] (NOT-SHOWN; per stratum covered-1 +0.0164 [+0.0111, +0.0210] / +0.0136 [+0.0049, +0.0217] (NOT-SHOWN), covered-2-4 +0.0029 [-0.0008, +0.0070] / -0.0002 [-0.0036, +0.0031] (NOT-SHOWN), covered-5+ -0.0011 [-0.0016, -0.0006] / +0.0004 [-0.0003, +0.0011] (NEGATIVE / NOT-SHOWN)), pooled +0.0198 [+0.0146, +0.0260] / +0.0270 [+0.0200, +0.0361] (NOT-SHOWN); the gain-to-margin ratio 1.0161 / 1.0464 of the stratified read's unseen-key margin (+0.1940 / +0.1741, reproduced here as +0.1940 [+0.1596, +0.2280] / +0.1741 [+0.1390, +0.2154]) — a ratio of the free arm's gain to the old margin, not a share of the margin closed. D_eq^R2 (history_tok_r2 - free_r2): on the unseen-key positions +0.0732 [+0.0623, +0.0834] / +0.0662 [+0.0559, +0.0773] (SIGNAL, no floor straddle; per stratum cur-unseen +0.0674 [+0.0558, +0.0791] / +0.0593 [+0.0489, +0.0716] (SIGNAL), prev-unseen +0.0720 [+0.0636, +0.0805] / +0.0681 [+0.0591, +0.0792] (SIGNAL), both-unseen +0.1006 [+0.0760, +0.1133] / +0.0882 [+0.0554, +0.1011] (SIGNAL)), pooled +0.0493 [+0.0425, +0.0577] / +0.0506 [+0.0433, +0.0594] (NOT-SHOWN / SIGNAL by the rule, the CI95 straddling the floor on both — the one label-bearing reading on which the harvests differ, a seed-level difference at the floor's edge on a cell the label does not read when the unseen-key gate is SIGNAL), on the covered positions (0.9013 / 0.8582 of the test positions) +0.0467 [+0.0402, +0.0547] / +0.0480 [+0.0411, +0.0564] (NOT-SHOWN; covered-1 +0.0586 [+0.0488, +0.0706] / +0.0598 [+0.0483, +0.0737] (SIGNAL), covered-5+ +0.0449 [+0.0385, +0.0525] / +0.0461 [+0.0398, +0.0536] (NOT-SHOWN)); the union D_add^R2 (history_bigram_r2 - free_r2) pooled +0.0549 [+0.0480, +0.0631] / +0.0556 [+0.0483, +0.0645] (SIGNAL, the CI95 containing the floor on both) and on the unseen keys +0.0764 [+0.0653, +0.0868] / +0.0696 [+0.0591, +0.0810] (SIGNAL); the twin's OWN address arm against the generalising free block (history_tok - free_r2, the reported co-reading) pooled +0.0411 [+0.0348, +0.0487] / +0.0390 [+0.0323, +0.0473] (NOT-SHOWN) and on the unseen keys -0.0031 [-0.0165, +0.0093] / -0.0081 [-0.0226, +0.0070] (NOT-SHOWN): without the smoothed current-token block no margin of the address arm over the generalising free arm is shown on the unseen keys — the surviving margin is the address block's over the free block when both carry the same encoding; what the smoothed current-token block adds to the address arm (history_tok_r2 - history_tok) +0.0083 [+0.0069, +0.0098] / +0.0116 [+0.0100, +0.0133] pooled and +0.0764 [+0.0715, +0.0814] / +0.0742 [+0.0689, +0.0801] on the unseen keys (no handicap: both CI-positive). The arithmetic on the unseen keys, exact in the paired deltas: the stratified read's located +0.1940 [+0.1596, +0.2280] / +0.1741 [+0.1390, +0.2154] equals the twin-arm co-reading plus the SHRINKAGE, and what remains against the generalising free block, +0.0732 [+0.0623, +0.0834] / +0.0662 [+0.0559, +0.0773], equals the co-reading plus what the smoothed block adds to the address arm; the rest, +0.1207 / +0.1080, is the free arm's gain net of the address arm's gain from the same encoding (the SHRINKAGE less history_tok_r2 - history_tok on the unseen keys), not a share of the margin closed. Q2, the stored final-norm state (clipped, z-scored and reduced to 1024 principal directions on train) in the free arm: D_eq^H (history_bigram_r2 - free_h, 3073 against 3073) +0.0111 [+0.0078, +0.0150] / +0.0106 [+0.0068, +0.0153] (NOT-SHOWN, CI-positive below the floor, no straddle), D_add^H (history_all_h - free_h) +0.0172 [+0.0139, +0.0210] / +0.0162 [+0.0126, +0.0206] (NOT-SHOWN, CI-positive below the floor, no straddle), the dilution anchor +0.0061 [+0.0057, +0.0064] / +0.0056 [+0.0052, +0.0060]; what H adds to the free arm (free_h - free_r2) +0.0438 [+0.0397, +0.0483] / +0.0450 [+0.0407, +0.0500]; h_only - free_tok -0.0051 [-0.0132, +0.0036] / +0.0035 [-0.0058, +0.0137] (NOT-SHOWN on both; the primary's point below zero, the replication's above); h_only - history_tok -0.1083 [-0.1129, -0.1039] / -0.1038 [-0.1090, -0.0992] (NEGATIVE: the pinned flag, h_only ABOVE history_tok, does not fire on either harvest). ARMS at B = 16, j = 1, pooled (primary / replication): free_tok 0.8110 / 0.8035, FREE_BIGRAM 0.8534 / 0.8449, history_tok 0.9143 / 0.9109, HISTORY_BIGRAM 0.9194 / 0.9155; free_r2 0.8732 / 0.8719, history_tok_r2 0.9225 / 0.9225, history_bigram_r2 0.9281 / 0.9275, free_backoff 0.8682 / 0.8645, free_replay 0.8644 / 0.8656, replay_cur 0.8216 / 0.8241, replay_zerofit 0.5675 / 0.5707, h_only 0.8059 / 0.8071, free_h 0.9169 / 0.9169, history_all_h 0.9342 / 0.9331; on the unseen-key positions FREE_BIGRAM 0.6018 / 0.6428, history_tok 0.7958 / 0.8169, free_r2 0.7989 / 0.8250, history_tok_r2 0.8721 / 0.8911, free_replay 0.8043 / 0.8282, h_only 0.7064 / 0.7459, free_h 0.8409 / 0.8715. The family's ends against the table: free_backoff - FREE_BIGRAM +0.0149 [+0.0109, +0.0196] / +0.0195 [+0.0141, +0.0268], free_replay - FREE_BIGRAM +0.0111 [+0.0051, +0.0181] / +0.0207 [+0.0129, +0.0308]; the selected k over each end: free_r2 - free_backoff +0.0050 [+0.0034, +0.0068] / +0.0074 [+0.0055, +0.0097], free_r2 - free_replay +0.0088 [+0.0075, +0.0099] / +0.0063 [+0.0049, +0.0076]; the replay alone: replay_cur - free_tok +0.0106 [+0.0046, +0.0177] / +0.0205 [+0.0129, +0.0303], replay_zerofit - free_tok -0.2435 [-0.2583, -0.2274] / -0.2328 [-0.2491, -0.2145]. Said in words (the tool's limit-arm flag reads ALL only): on the unseen-key positions the pure replay reads above the selected k — free_r2 - free_replay -0.0054 [-0.0081, -0.0026] / -0.0032 [-0.0066, +0.0001] (NEGATIVE / NOT-SHOWN). REPLAY AGREEMENT on test (no fit): REPLAY(tokens[t-1]) against S[t] pooled 0.4003 / 0.4033, at gate 0 0.2403 / 0.2416 (the lowest gate on both), the maximum 0.5096 / 0.5150 at gate 10, gate 15 0.3147 / 0.3188; chance 8 / 64 = 0.125. The embedding-only replay is a weak proxy at gate 0, where F-26's GATE-0 NOTE records two unresolved readings (routing on the previous token through layer 0's attention, or the stored gate-0 slot lagging one forward) and a low embedding-only agreement fits both. HIDDEN BLOCK: 292,946 / 147,676 train positions, kept-variance share 0.7952 / 0.7970, the 1,024th and 1,025th eigenvalues 0.6015 / 0.6017 and 0.6011 / 0.6016 (a near tie at the cut: a different BLAS could swap the last directions, which touches the H arms only — and the H arms have no earlier rows to reproduce, so REPRODUCE cannot cover them). THE PINNED SENTENCES, in their words: Q1 — "against a free encoding that is never at the marginal (the model's own routers replayed on the token embedding as the prior of the smoothed count tables, k = 128), the address block still adds a material margin over the equal-dims free block on the unseen-key positions (+0.0732 [+0.0623, +0.0834] / +0.0662 [+0.0559, +0.0773], SIGNAL on both; pooled +0.0493 [+0.0425, +0.0577] / +0.0506 [+0.0433, +0.0594], NOT-SHOWN / SIGNAL by the rule with the CI95 straddling the +0.05 floor on both — said in words), while the encoding recovers +0.1971 [+0.1697, +0.2266] / +0.1822 [+0.1557, +0.2167] of hit rate there (the gain-to-margin ratio 1.0161 / 1.0464): the located margin is NOT shown to be table sparsity at this encoding and the coverage caveat stands as a scoped limitation, not a retraction, at this reader (B = 16, j = 1)"; Q2 — "with the hidden block in the free arm, no material margin of the address block over the equal-dims free block (+0.0111 [+0.0078, +0.0150] / +0.0106 [+0.0068, +0.0153]) and none of the address stream over the free bigram plus the hidden block (+0.0172 [+0.0139, +0.0210] / +0.0162 [+0.0126, +0.0206]) is shown: at this reader the forward's own final-norm state (at 1024 of 2048 directions, keeping 0.7952 / 0.7970 of the train variance) plus two tokens' smoothed routing profiles show no material address margin (NON-CRITERION; K-2 is untouched — this is a prefetch hit-rate reader, not the K-2 reader); stage 2's address-beyond-free motivation is struck (PREFETCH-HEADROOM, which does not depend on this margin, stands)". THE STRIKE APPLIES by the pinned condition: neither Q2 CI95 contains the floor on either harvest, the replication agrees on Q2, and the READER GUARD did not fire on either (free_h - free_r2 +0.0438 [+0.0397, +0.0483] / +0.0450 [+0.0407, +0.0500], the dilution anchor +0.0061 [+0.0057, +0.0064] / +0.0056 [+0.0052, +0.0060], both CI-positive); beside the strike, as pinned: D_add^H on the unseen-key positions +0.0399 [+0.0279, +0.0514] / +0.0286 [+0.0195, +0.0386] (NOT-SHOWN; the primary's CI95 containing the floor), and D_eq^H there +0.0343 [+0.0213, +0.0465] / +0.0231 [+0.0131, +0.0342] (NOT-SHOWN, no straddle); on the both-unseen stratum (672 / 422 positions, a stratum that draws 0.3482 / 0.5379 of its positions from one test prompt, prompt 13, by the stratified read) D_eq^H +0.0859 [+0.0340, +0.1002] / +0.0721 [+0.0261, +0.0926] and D_add^H +0.0868 [+0.0424, +0.1001] / +0.0709 [+0.0299, +0.0895] read SIGNAL / SIGNAL by the rule with every CI95 containing the floor — said beside the strike as the registration requires; the strike reads the pooled cells only. LAST RUNG, as pre-committed: R2 closes the free-baseline ladder under this finding (one-token table -> free bigram -> stratified read -> generalising baseline with the hidden arm); no further free-baseline rung is registered on this margin; the next move is the write-ups or stage 2 under its own registration — and with its address-beyond-free motivation struck, stage 2 would need a motivation of its own; Q1's survival is no such motivation (a reroute around a fired strike). SCOPING NOTE (2026-10-03, a dated note, not a rewrite): LAST RUNG's pinned SURVIVES wording, 'against the free arms this registration builds', is NOT attached to Q1 — Q1's partner is free_r2 alone, and against free_h, the strongest free arm this registration builds, Q2 reads H-ABSORBS; Q1's label reads 'against the generalising free block', nothing wider. Said in words beside the labels: the pooled Q1 reading +0.0493 [+0.0425, +0.0577] / +0.0506 [+0.0433, +0.0594] reads NOT-SHOWN / SIGNAL by the rule with the CI95 straddling the floor on both — the one label-bearing reading on which the harvests differ, said in words; the three reported-arm grid edges; the B = 8 / 32 disagreements above; D_add^H's unseen-key CI containing the floor on the primary; the both-unseen Q2 cells SIGNAL with the floor inside every CI; h_only below history_tok on both. NON-CRITERION: F-26's labels, the twin's SURVIVES and the stratified read's labels stand as computed; F-26's sentence "nothing in this landing reads ADDRESS-BEYOND-TOKEN as context information in the address stream beyond free information or as support for the sideband premise" STANDS; no kill status or regime moves; no wall-clock, bytes or latency sentence (D-9; K-5 / K-6 stay UNTESTED). The coverage caveat's standing after this rung: on the unseen-key positions the address block's margin over the generalising free block is material and the located margin is NOT shown to be table sparsity at this encoding (Q1), while no material margin (at or above +0.05) of the address block over the equal-dims free block, nor of the address stream over the free bigram plus the hidden block, is shown (both readings CI-positive below the floor) once the free arm holds the forward's own final-norm state beside the smoothed bigram (Q2) — the first is a reading against free_r2, the second against the strongest free arm this registration builds; both are hit-rate readings at linear readers on these harvests. CLEAN-CLONE RERUN (the recompute critic's named check, run after its GO): the primary rerun from a fresh clone of the freeze commit (core.autocrlf=false; the harvest directory pointed at the main tree's shards; the stratified payload the clone's own; the replay from the cache) to a scratch path, 673.5 s of elapsed time on the same box and BLAS (scipy-openblas 0.3.30): BYTE-IDENTICAL: the state file (sha 7d43c9678090), the conclusion, the k and lambda val tables, the selections, the hidden block's spectrum, the replay probe and the replay agreement all reproduce exactly — so the thinnest of the k and lambda selections in the primary run, the k selection's margin, and the val tables the committed rows cannot recompute are reproducible on this box. Both predictions scored in the block's OUTCOME; critic recompute: GO at maximum absolute discrepancy 0.0 over 74,366 values (37,183 per harvest), the record in the block's OUTCOME.)
    sideband/FINDINGS.md:297 · d147e8c0cd6b
  40. 2026-10-02sidebandF-26amendment

    status (2026-10-02; STRATIFIED COVERAGE READ, the post-hoc diagnostic registered as the dated amendment "OQ-9 STAGE-1 STRATIFIED COVERAGE READ REGISTRATION" in docs/BLUEPRINT.md before any readout of it ran, both runs from the clean freeze commit fea2431; …

    The full line, as written
    - status (2026-10-02; STRATIFIED COVERAGE READ, the post-hoc diagnostic registered as the dated amendment "OQ-9 STAGE-1 STRATIFIED COVERAGE READ REGISTRATION" in docs/BLUEPRINT.md before any readout of it ran, both runs from the clean freeze commit fea2431; evidence results/real/prefetch_strat_s3.json (primary) and results/real/prefetch_strat_s3_r.json (replication), each with .meta.json and .state.jsonl; instrument tools/run_prefetch_strat.py, which imports the frozen twin tools/run_prefetch_bigram.py and refuses a modified one; tests/test_prefetch_strat.py, tests/test_prefetch_strat_landing.py): LABEL FALLBACK-LOCATED / PT-SURVIVES on the primary; the replication AGREES on both questions (a SEED replication on the same prompts; the primary's label is the label). PRECONDITION REPRODUCE EXACT on both: the stratified rows summed over the six coverage strata reproduce every one of the twin's 12,960 committed state rows to the integer, and the pooled D_eq equals the twin's committed D_eq (+0.0609 [+0.0514, +0.0712] / +0.0659 [+0.0547, +0.0788]), CI included. Both gates STRADDLE the +0.05 floor on both harvests, so the pinned straddle sentences apply and neither branch is claimed. On the twin's bigram positions at B = 16, j = 1, PRIMARY / REPLICATION — Q1 (COVERAGE): on the COVERED positions (both token keys seen in train; 0.9013 / 0.8582 of the test positions) D_eq^COV +0.0463 [+0.0398, +0.0541] / +0.0481 [+0.0411, +0.0564] (NOT-SHOWN / NOT-SHOWN: CI-positive below the floor, the CI95 containing it); on the FALLBACK positions (at least one key unseen; 0.0987 / 0.1418) D_eq^FB +0.1940 [+0.1596, +0.2280] / +0.1741 [+0.1390, +0.2154] (SIGNAL / SIGNAL); per unseen stratum, current key unseen (0.0436 / 0.0636) +0.1861 [+0.1662, +0.2033] / +0.1736 [+0.1533, +0.1936], previous key unseen (0.0438 / 0.0641) +0.1209 [+0.1071, +0.1340] / +0.1113 [+0.0979, +0.1258], both unseen (0.0113 / 0.0141) +0.5068 [+0.4069, +0.5300] / +0.4613 [+0.3354, +0.5115]; PT GAIN on the unseen-key positions +0.0805 [+0.0657, +0.0959] / +0.0805 [+0.0572, +0.1106]. Q2 (p_t), with the forward's stored output-distribution summaries (mass16 ++ mass64 ++ entropy / ln V, 81 coordinates) in BOTH arms, on ALL positions: D_eq^PT = history_tok_pt - free_bigram_pt +0.0532 [+0.0455, +0.0619] / +0.0565 [+0.0480, +0.0664] (SIGNAL / SIGNAL, the CI95 containing the floor), D_add^PT = history_bigram_pt - free_bigram_pt +0.0583 [+0.0506, +0.0669] / +0.0609 [+0.0525, +0.0709] (SIGNAL / SIGNAL), the dilution anchor history_bigram_pt - history_tok_pt +0.0051 [+0.0048, +0.0054] / +0.0044 [+0.0041, +0.0047]; PT GAIN on ALL +0.0089 [+0.0066, +0.0118] / +0.0121 [+0.0077, +0.0183] (NOT-SHOWN / NOT-SHOWN), on COVERED +0.0011 [+0.0008, +0.0014] / +0.0008 [+0.0004, +0.0011], on the current-key-unseen positions +0.1118 [+0.0956, +0.1276] / +0.1110 [+0.0930, +0.1277]; the asymmetric replacement read D_pt = history_tok - free_bigram_pt +0.0520 [+0.0442, +0.0608] / +0.0539 [+0.0455, +0.0636] (reported); pt_only 0.6032 / 0.6066, free_bigram_pt 0.8623 / 0.8570, history_tok_pt 0.9155 / 0.9135, history_bigram_pt 0.9205 / 0.9179 against history_tok 0.9143 / 0.9109 and FREE_BIGRAM 0.8534 / 0.8449. Q2 BY SCOPE (reported): on COVERED D_eq^PT +0.0455 [+0.0390, +0.0532] / +0.0475 [+0.0407, +0.0557] (NOT-SHOWN / NOT-SHOWN, the floor inside the CI) and D_add^PT +0.0512 [+0.0447, +0.0589] / +0.0525 [+0.0457, +0.0608] (SIGNAL / SIGNAL, the floor inside the CI); on FALLBACK D_eq^PT +0.1231 [+0.1058, +0.1382] / +0.1110 [+0.0950, +0.1261] and D_add^PT +0.1229 [+0.1060, +0.1371] / +0.1119 [+0.0960, +0.1271] — on COVERED D_eq^PT reads NOT-SHOWN / NOT-SHOWN and D_add^PT SIGNAL / SIGNAL by the rule, both with the floor inside the CI (the twin's D_add on covered likewise SIGNAL / SIGNAL, +0.0520 / +0.0531); the equal-dims margin is indicated on the unseen-key positions, while the union's margin is not confined there (D_add^PT reads SIGNAL on COVERED as well); PT-SURVIVES inherits the coverage caveat and is never read as 'the address block beats free p_t information' on the covered positions. IN THE PINNED WORDS (the straddle variants, both harvests): Q1 — "the address block's equal-dims margin on the positions whose token keys the train tables saw reads NOT-SHOWN by the rule with a CI95 that straddles the +0.05 floor, while a material margin is shown on the positions with at least one unseen key (the current-token table blind in both arms, the free previous-token table blind, or both): the location is indicated, not settled, at this power, at this reader (B = 16, j = 1)"; Q2 — "with the forward's own output-distribution summaries (mass16, mass64, entropy) in both arms, both Q2 readings are SIGNAL by the rule while at least one CI95 straddles the +0.05 floor: the twin's verdict with p_t is not settled at this power, at this reader (B = 16, j = 1)". So the coverage caveat is neither retired nor confirmed at this power: its location is INDICATED on the unseen-key positions, where at least one free table is blind and the address block's margin is large, while on the covered positions the equal-dims margin sits just below the floor with the floor inside its CI. By the point identity the covered positions carry +0.0417 / +0.0412 of the pooled +0.0609 / +0.0659 (0.686 / 0.626 of the margin), so 'indicated on the unseen-key positions' is a statement about per-position materiality, not about where most of the pooled margin accrues: most of it accrues on the covered positions, below the floor per position. And with p_t's summaries in the baseline on both sides the twin's two readings both read SIGNAL, the equal-dims one with the floor inside its CI, so the twin's verdict with p_t is not settled at this power either. SAID IN WORDS, as pinned: the pt_only arm's lambda sits at the grid's lower edge (0.1) on both harvests (a reported-only arm; its 81-coordinate design is the smallest and its hit rate is far below free_tok); the covered split by the smaller key's train occurrences — one occurrence (0.0529 / 0.0665 of positions) D_eq +0.0658 [+0.0573, +0.0750] / +0.0683 [+0.0579, +0.0808] (SIGNAL / SIGNAL), two to four (0.0976 / 0.1346) +0.0552 [+0.0472, +0.0645] / +0.0503 [+0.0417, +0.0610] (SIGNAL / SIGNAL), five or more (0.7508 / 0.6571) +0.0438 [+0.0374, +0.0514] / +0.0456 [+0.0390, +0.0536] (NOT-SHOWN / NOT-SHOWN) — reported readouts whose readings differ from the covered union's: the union's reading is the well-covered keys', and the sparse-key strata read at or above the floor; the union stands as the gate; the both-unseen stratum (0.0113 / 0.0141; 27 / 29 of the 30 prompts) draws 0.3482 / 0.5379 of its positions from one test prompt (prompt 13); without that prompt the stratum reads +0.4907 / +0.3672 (the critic's unregistered leave-one-prompt-out, scratch, not evidence); the reported budgets DISAGREE with the label budget's readings and are said so: at B = 8 D_eq^COV +0.0642 [+0.0585, +0.0708] / +0.0672 [+0.0612, +0.0747] (SIGNAL / SIGNAL, no straddle), where Q1 would read COVERED-MATERIAL, and at B = 32 D_eq^COV +0.0136 [+0.0109, +0.0169] / +0.0145 [+0.0115, +0.0182] with D_eq^PT +0.0184 [+0.0147, +0.0225] / +0.0200 [+0.0159, +0.0251] and D_add^PT +0.0197 [+0.0160, +0.0239] / +0.0213 [+0.0171, +0.0264] all CI-positive below the floor, where Q2 would read PT-ABSORBS — on positions where the twin's own pooled margin was already CI-positive below the floor (+0.0261 / +0.0287; no material margin is shown there); the stage-2 strike is pinned to the label budget, which reads PT-SURVIVES, so no stage-2 sentence follows; where the margin sits is budget-dependent, and 'indicated, not settled' is the honest wording. At B = 16, D_eq^PT and D_pt agree on every scope's reading on both harvests; both match D_eq's per-scope readings except on the replication's two-to-four stratum, where D_eq reads SIGNAL (+0.0503) and D_eq^PT / D_pt read NOT-SHOWN (+0.0489 / +0.0484). 'The replication agrees' is seed robustness only: both harvests read the same 30 test prompts, and prompt 13 is the largest contributor to both-unseen on both. NON-CRITERION: F-26's labels and the twin's SURVIVES stand as computed; F-26's sentence 'nothing in this landing reads ADDRESS-BEYOND-TOKEN as context information in the address stream beyond free information or as support for the sideband premise' STANDS; no kill status or regime moves; no wall-clock, bytes or latency sentence (D-9; K-5 / K-6 stay UNTESTED). NEXT RUNG (unrun, unregistered): a free baseline whose encoding handles unseen and sparse keys (smoothed, hashed or embedding tables, the R2 of the plan of record), so the margin is read against a free arm that is never at the marginal and never a single train sample; and F-24's full-free block ((a) ++ TOK ++ z-scored h_a, stored in these harvests, F-23) as the stronger free arm for the p_t question — PT-SURVIVES says nothing about it. Runtimes 175.7 / 91.9 s of elapsed time; 110,448 / 111,696 state rows; critic recompute: see the block's OUTCOME.)
    sideband/FINDINGS.md:296 · d147e8c0cd6b
  41. 2026-10-01sidebandF-26amendment

    status (2026-10-01; FREE-BIGRAM TWIN, the post-hoc diagnostic registered as the dated amendment "OQ-9 STAGE-1 FREE-BIGRAM TWIN REGISTRATION" in docs/BLUEPRINT.md before any readout of it ran, both runs from the clean freeze commit d55e62f; …

    The full line, as written
    - status (2026-10-01; FREE-BIGRAM TWIN, the post-hoc diagnostic registered as the dated amendment "OQ-9 STAGE-1 FREE-BIGRAM TWIN REGISTRATION" in docs/BLUEPRINT.md before any readout of it ran, both runs from the clean freeze commit d55e62f; evidence results/real/prefetch_bigram_s3.json (primary) and results/real/prefetch_bigram_s3_r.json (replication), each with .meta.json and .state.jsonl; instrument tools/run_prefetch_bigram.py, which imports the frozen stage-1 instrument and refuses a modified one; tests/test_prefetch_bigram.py, tests/test_prefetch_bigram_landing.py): LABEL SURVIVES on the primary; the replication AGREES on the label and on both readings (a SEED replication on the same prompts; the primary's label is the label). On the bigram positions (t >= 1) at B = 16, j = 1 — PRIMARY: re-measured ADDRESS-BEYOND-TOKEN +0.1033 [+0.0927, +0.1146] against the committed +0.1032 (gap +0.0001; precondition HELD); FREE_BIGRAM = free_bigram_b (encoding prev_table) 0.8534 against free_tok 0.8110 and history_tok 0.9143; HISTORY_BIGRAM = history_bigram_b 0.9194; BIGRAM GAIN +0.0424 [+0.0407, +0.0441] (NOT-SHOWN); D_eq = history_tok - FREE_BIGRAM +0.0609 [+0.0514, +0.0712] (SIGNAL); D_add = HISTORY_BIGRAM - FREE_BIGRAM +0.0661 [+0.0569, +0.0762] (SIGNAL); the free previous token accounts for 0.4104 of the re-measured margin (point ratio); largest per-gate gain at gate 0 (+0.2201; next gate 1 +0.0890); gate 0 gain +0.2201, share of the summed gain 0.3245, cap 0.0203; per-gate gains min +0.0130 max +0.2201; REPORTED BUDGETS, in words as pinned: at B = 8 D_eq +0.0755 [+0.0677, +0.0838] (SIGNAL) and D_add +0.0837 [+0.0762, +0.0918] (SIGNAL); at B = 32 D_eq +0.0261 [+0.0194, +0.0336] and D_add +0.0275 [+0.0208, +0.0349] are CI-positive BELOW the floor (NOT-SHOWN by the rule, both; the label would read FREE-HISTORY at B = 32), as F-26's own margin was not material at B = 32 (+0.0405 / +0.0434). REPLICATION: re-measured ADDRESS-BEYOND-TOKEN +0.1074 [+0.0953, +0.1207] against the committed +0.1072 (gap +0.0002; precondition HELD); FREE_BIGRAM = free_bigram_b (encoding prev_table) 0.8449 against free_tok 0.8035 and history_tok 0.9109; HISTORY_BIGRAM = history_bigram_b 0.9155; BIGRAM GAIN +0.0414 [+0.0398, +0.0430] (NOT-SHOWN); D_eq = history_tok - FREE_BIGRAM +0.0659 [+0.0547, +0.0788] (SIGNAL); D_add = HISTORY_BIGRAM - FREE_BIGRAM +0.0705 [+0.0592, +0.0834] (SIGNAL); the free previous token accounts for 0.3858 of the re-measured margin (point ratio); largest per-gate gain at gate 0 (+0.2161; next gate 1 +0.0860); gate 0 gain +0.2161, share of the summed gain 0.3261, cap 0.0206; per-gate gains min +0.0111 max +0.2161; REPORTED BUDGETS, in words as pinned: at B = 8 D_eq +0.0817 [+0.0722, +0.0926] (SIGNAL) and D_add +0.0886 [+0.0793, +0.0993] (SIGNAL); at B = 32 D_eq +0.0287 [+0.0208, +0.0390] and D_add +0.0300 [+0.0221, +0.0404] are CI-positive BELOW the floor (NOT-SHOWN by the rule, both; the label would read FREE-HISTORY at B = 32), as F-26's own margin was not material at B = 32 (+0.0405 / +0.0434). PINNED SENTENCE: "the address block adds a material margin over an equal-dims free bigram block and the address stream adds a material margin over the free bigram at this reader (B = 16, j = 1)". IN THE PINNED WORDS: the address block adds +0.0609 [+0.0514, +0.0712] over an equal-dims free bigram block and +0.0661 [+0.0569, +0.0762] over the free bigram at this reader; the critic's bigram objection is answered at THIS ENCODING (two additive 1024-dim train tables), not for previous-token information in general. COVERAGE CAVEAT (a caveat on the interpretation; the mechanical label is not contested): on the test bigram positions the current-token key has no train count at a share 0.0549 (previous-token key 0.0551; either 0.0987; both 0.0113; token-only counts). F-26's HARD BOUND (D_eq - f) / (1 - f), which needs only that D_eq at a fallback position is at most 1, gives the D_eq on positions where BOTH keys are covered at least -0.0419 here (f the either-unseen share; with the current-token share alone, +0.0063), so the bound does NOT establish a material margin on covered keys. The critic's UNREGISTERED val-only stratified diagnostic at this landing (readers re-fit exactly as the instrument fits them and read at the payload's selected lambdas on the VAL bigram positions; pooled val reproduces the payload's val table to six decimals; test rows not read; re-run by the assistant with identical output; scratch only, NOT evidence) reads D_eq +0.0454 where both keys have a train count (0.9115 of val) — BELOW the floor — against +0.1844 where the current token is unseen, +0.1190 where the previous token is unseen and +0.4491 where both are; D_add +0.0511 on the covered positions. On the TEST rows (state-row arithmetic joined to token-only coverage): per-prompt D_eq tracks the per-prompt fallback share (r 0.787 with the current-token share, 0.817 with either key unseen); the fifteen prompts at or below the median share pool to D_eq +0.0429 against +0.0793 above it; D_eq is at or above the floor on 16 of 30 prompts (positive on 30; range +0.0313..+0.1431; F-26's margin cleared the floor on 30 of 30). READING: the material part of the surviving pooled margin sits where the free tables are blind — positions whose token key the train tables never saw: every fallback stratum on val reads D_eq +0.1190..+0.4491 against +0.0454 on covered positions. The diagnostic LOCATES the margin; it does not identify the mechanism. At unseen PREVIOUS tokens the location is consistent with the registered identity qualification (S[t] is the routing of the forward tokens[t-1] fed); at unseen CURRENT tokens — the largest stratum — S[t] was computed before tokens[t] was sampled and cannot expose that token's identity: it can carry only the forward's PREDICTION of the token (information the output distribution p_t holds, free under D-5; the harvest's stored mass16 / mass64 / entropy summaries of p_t were never in the prefetch free baseline — stage 1 read them only in the reported-only front-loading arm) or context. SURVIVES therefore holds at THIS ENCODING on the pooled positions as registered, and F-26's sentence 'nothing in this landing reads ADDRESS-BEYOND-TOKEN as context information in the address stream beyond free information or as support for the sideband premise' STANDS unchanged. NEXT RUNG (unrun, unregistered; this CORRECTS the rung the registration named): the STRATIFIED covered-against-fallback read of D_eq on the TEST split at the selected lambdas, with a free arm built from the harvest's mass16 / mass64 / entropy beside it (the critic's cheapest experiment: Bigram.fit plus a per-position mask, about 75 / 45 s of elapsed time on val) — its branches are pinned at that rung's own registration, not here; in outline, covered-position D_eq SIGNAL on the primary would retire this caveat and NOT-SHOWN would make its materiality a coverage effect at this encoding; the pair-keyed (true bigram) table named at registration has WORSE coverage and would move D_eq by this very location effect, so it is not the next rung; a longer free history (tokens[t-2]) shares the problem unless its encoding handles unseen keys; on the replication the same diagnostic reads covered-position D_eq +0.0472 (0.8634 of val), the hard bound -0.0128, r 0.804, the floor cleared on 17 of 30 prompts. TWIN-EXEMPTION ANCHORS: free_all - FREE_BIGRAM +0.0017 [+0.0015, +0.0020] and HISTORY_BIGRAM - history_tok +0.0052 [+0.0049, +0.0054], CI-clear positive — above the priced range's upper end (-0.0045..+3e-5) by about two orders of magnitude, one order below the floor rather than two (F-24's corrected twin helped the same way); the readings are unaffected (D_add - D_eq equals the second anchor exactly), the price was wrong; replication +0.0016 [+0.0014, +0.0019] and +0.0046 [+0.0043, +0.0049]. GATE-0 NOTE (the critic's; an unresolved alternative, not a finding): the statistics read here put gate 0 one forward behind gates 1-15 — its best key is tokens[t-1] where the others' is tokens[t]; in the stage-1 payload the intra-forward reader at gate 1, which reads gate 0 of the same stored forward, scores 0.6676, below the one-token table at gate 1 (0.7658) and below the previous forward's full routing (0.7422) — which 'gate 0 routes on the previous token' and 'the stored gate-0 slot lags one forward' predict equally well; neither leaks future information, gate 0's own D_eq is +0.0014 (its stage-1 margin +0.2208 almost fully absorbed), so the surviving D_eq sits at gates 1-15 and the label is unaffected; gate 0 contributes 0.0138 of the pooled gain +0.0424; replication gate 0 D_eq +0.0084, contribution 0.0135. NON-CRITERION: the mechanical labels of this finding (stage label, four readings) were computed on the registered arms and stand as computed; this line changes nothing in them and adds the standing of the sideband sentence at this encoding: the margin over the one-token table survives the free previous token at equal dims on the pooled positions, with the coverage caveat; the finding's 'nothing reads this as context or as support for the sideband premise' stands; D_add is never read as context; no wall-clock, bytes or latency sentence follows (D-9; K-5 / K-6 stay UNTESTED; runtimes are elapsed time of the instrument process (time.time), not latency numbers). PREDICTIONS SCORED: HEADLINE: both pinned modal expectations put D_eq below material; it read SIGNAL on the pooled positions, with the CI lower +0.0014 / +0.0047 above the floor, and below the floor on covered positions in the unregistered diagnostic (COVERAGE CAVEAT). THE CRITIC'S (2026-09-15): 'the margin drops to or below the +0.05 floor' -> D_eq +0.0609 [+0.0514, +0.0712] / +0.0659 [+0.0547, +0.0788] (SIGNAL / SIGNAL): FAILED / FAILED; 'driven by gate 0' -> argmax gate 0 / 0: HELD / HELD under the pinned argmax rule (a near-certain sub-clause given the committed per-gate margins, so it carries little weight); gate 0 contributes 0.0138 / 0.0135 of the pooled gain +0.0424 / +0.0414 against its cap 0.0203 / 0.0206 (share of the summed gain 0.3245 / 0.3261): the two halves of that prediction were close to incompatible — a drop to the floor needed at least 0.053 of pooled gain and gate 0 could supply at most 0.020 — an error the registration recorded before the run; the reading that gate 0 alone could not have carried a drop to the floor is correct. THE ASSISTANT'S (2026-10-01): (1) encoding (b) prev_table -> prev_table / prev_table: HELD / HELD; (2) FREE_BIGRAM 0.84..0.88 -> 0.8534 / 0.8449: HELD / HELD; (3) BIGRAM GAIN +0.03..+0.07 -> +0.0424 [+0.0407, +0.0441] / +0.0414 [+0.0398, +0.0430]: HELD / HELD; (4) gate 0 the largest gain, at least twice the next gate's and at least a quarter of the summed gain -> HELD (gate 0 +0.2201, 2.47x gate 1 +0.0890; share 0.3245) / HELD (gate 0 +0.2161, 2.51x gate 1 +0.0860; share 0.3261); (5) each of gates 1-15 between 0 and +0.04 -> FAILED / FAILED: 3 / 3 of 15 gates exceed +0.04 (gates 1, 2, 4 / 1, 2, 4; range +0.0130..+0.0890 / +0.0111..+0.0860); (6) D_eq +0.03..+0.08, NOT-SHOWN or a SIGNAL at the floor's edge -> +0.0609 [+0.0514, +0.0712] / +0.0659 [+0.0547, +0.0788] (SIGNAL / SIGNAL): HELD / HELD on the range; 'at the floor's edge' was never quantified — the CI lower sits +0.0014 / +0.0047 above the floor; (7) D_add +0.05..+0.09 SIGNAL -> +0.0661 [+0.0569, +0.0762] / +0.0705 [+0.0592, +0.0834] (SIGNAL / SIGNAL): HELD / HELD; (8) label COMPLEMENTARY most likely, SURVIVES the alternative -> SURVIVES / SURVIVES: MODAL MISSED; the named alternative occurred (not rounded to held — the F-26 'MISSED-HIGH, not rounded' precedent); (9) replication agrees on the label -> HELD; (10) precondition gap under 0.003 -> +0.0001 / +0.0002: HELD / HELD; (11) every lambda interior -> grid edges none / none: HELD / HELD. WOULD-SURPRISE LIST (primary / replication): D_eq at or above +0.10: no / no; D_eq NEGATIVE: no / no; D_add NOT-SHOWN: no / no; encoding (a) selected: no / no; largest gain not at gate 0: no / no; precondition gap above 0.005: no / no. CRITIC RECOMPUTE (2026-10-01): the sideband-critic recomputed every table (27 per harvest, pooled and per-gate, with CIs), every delta (30 per harvest) and every label field from the two state files with its own parser and bootstrap, re-deriving the FREE_BIGRAM encoding and all nine lambdas from the val tables: max abs discrepancy 0.0 over 1,993 numeric values on each harvest; the tool's --conclude-from exits 0 on both; 446 numbers quoted in the landing texts checked against the payloads and metas with 82 named-cell bindings, 0 misquotes; the sha bindings (instrument working copy = HEAD = both payloads = the landing test's pin; the frozen parent at 8529532d4d78; the twelve shard metas = the state headers = the bytes on disk; the committed cells on disk and at HEAD), the split (30 test prompts), the pairing (per-(prompt, gate) trials identical across all nine arms at every budget), the position counts (one anchor per row dropped, equal to the committed j = 2 counts) and the freeze-to-run timing (freeze 13:34:00, primary 13:34:18 to 13:38:37, replication 13:38:38 to 13:41:01, every landing edit after 13:42) confirmed; leakage none found. THREE BLOCKING objections, applied as one batch: a guard test that went red once the payload existed (fixed in the test, never in the instrument); the pinned B = 32 reading not said in words (now said, with B = 8); the COVERAGE CAVEAT above, which the SURVIVES texts had shipped without (now carried on every text with the hard bound, the critic's unregistered diagnostic named as such and re-run by the assistant with identical output, and the next rung corrected). Advisories applied: the prediction scoring's units (gate 0's contribution against its cap, not the share), '3 of 15 gates', the CI-lower distances, 'MODAL MISSED', the plain headline; the twin-exemption anchors outside their priced range; the gate-0 alignment alternative recorded; the landing test's comment, fail-on-missing and gate-0 pin; the count words in CLAUDE.md and SHARP_EDGES; 'elapsed time' for the runtime. Nothing declined. SECOND PASS (the same day, a scoped text check of the batch): 659 numbers and 122 named-cell bindings checked, 0 misquotes; the three blockers confirmed closed; every 'context' / 'premise' occurrence a negation; ONE BLOCKING objection the batch had introduced, applied: the caveat's READING asserted a mechanism — an identity channel — that the diagnostic never measured, and misapplied it at unseen CURRENT tokens, where S[t] predates the token and can carry only the forward's prediction of it (free as p_t) or context; the READING now locates the margin without naming its mechanism, and the next rung gains a free arm from the stored p_t summaries. Advisories applied: the 'can only inflate D_eq' deduction replaced by the measured strata; the hard bound recomputed for the both-covered set (either-unseen share); 'coverage artefact' replaced by 'materiality a coverage effect' with the branch deferred to that rung's registration; 'opposite sign' replaced by 'above the priced range'; 'every statistic' by 'the statistics read here'; the replication's duplicated caveat paragraph trimmed to its numbers; the two remaining 'CPU time' phrases. Nothing declined. Full outcome: the OUTCOME paragraph of the amendment block in docs/BLUEPRINT.md.
    sideband/FINDINGS.md:295 · d147e8c0cd6b

September 202627 entries

  1. 2026-09-15sidebandfindings table row F-26other

    | F-26 | 2026-09-15 | OQ-9 STAGE 1 (the prefetch half; NON-CRITERION, a K-6 datum): on the OLMoE S3 harvests the router address stream is a prefetch source with headroom over the recency cache (HISTORY +0.2867 / +0.2901) and a material margin over the free …

    The full line, as written
    | F-26 | 2026-09-15 | OQ-9 STAGE 1 (the prefetch half; NON-CRITERION, a K-6 datum): on the OLMoE S3 harvests the router address stream is a prefetch source with headroom over the recency cache (HISTORY +0.2867 / +0.2901) and a material margin over the free input token alone (ADDRESS-BEYOND-TOKEN +0.1032 [+0.0926, +0.1146] / +0.1072 [+0.0952, +0.1206] at B = 16 of 64, j = 1; below the +0.05 floor at B = 32, j = 1); INTRA-FORWARD +0.4861 / +0.4816; DRAFT ceiling (horizon 2) +0.1556 / +0.1545; stage PREFETCH-HEADROOM, the seed replication agrees on every label | stands (NON-CRITERION; K-6 stays UNTESTED; prior art occupies every cell); DATED STATUS 2026-10-01: the equal-dims free-bigram twin reads SURVIVES at that encoding (D_eq +0.0609 [+0.0514, +0.0712] / +0.0659 [+0.0547, +0.0788]; D_add +0.0661 / +0.0705; below the floor at B = 32, j = 1) with the coverage caveat (covered-key D_eq +0.0454 / +0.0472 on val in an unregistered diagnostic), read neither as context nor as support for the sideband premise; the stratified test-split read is the named, unregistered next rung; DATED STATUS 2026-10-16 (derived-ledger levelling, image audit 2026-10-15 section B item 6; the entry's status lines are the record): the stratified test-split read ran on 2026-10-02 as a registered post-hoc diagnostic (FALLBACK-LOCATED / PT-SURVIVES), the generalising free baseline R2 on 2026-10-03 (R2-FALLBACK-SURVIVES / H-ABSORBS, stage 2's address-beyond-free motivation STRUCK) and the horizon read on 2026-10-04 (LEAD-TIME-ABSORBED; OQ-9's prefetch half closes at this reader); the ladder is closed |
    sideband/FINDINGS.md:63 · d147e8c0cd6b
  2. 2026-09-14sidebandfindings table row F-25amendment

    | F-25 | 2026-09-14 | [read under the 2026-09-07 rider amendment after the registered rider fired at S1 L3 (raw leg 2 +0.0132 against the 0.01 band); diagnostic branch AMENDABLE, mechanism TWIN-MATCHED; a post-hoc reroute around a fired kill condition: …

    The full line, as written
    | F-25 | 2026-09-14 | [read under the 2026-09-07 rider amendment after the registered rider fired at S1 L3 (raw leg 2 +0.0132 against the 0.01 band); diagnostic branch AMENDABLE, mechanism TWIN-MATCHED; a post-hoc reroute around a fired kill condition: the leg that fired was replaced after the fire by a twin-referenced dependence-excess leg whose S1 seed-0 values at both spaces were known from the diagnostic payload when the amendment was written; it is a fresh test only at the other three harvests; the E leg sees every non-causal block the raw leg sees (the twin lacks it) and is sharper under a drifted reader; what it drops (prediction-marginal drift and twin-shared position information) are non-leaks; S1 diagnostic X_raw at seed 0 +0.0006, the registered statistic, twin-referenced, would also have passed] OQ-18 COARSE-LABEL FULL-FREE RE-TEST at L2 / L3, both sources, both harvests: S1 L2 ABSORBED; S1 L3 NOT-SHOWN; S3 L2 ABSORBED; S3 L3 SPLIT (every L3 label read at the L3 group lens (S1: one mega-group of 455/510 concepts including all 165 zero-degree concepts, the next group 10; S3: two mega-groups of 268 and 193 = 461/510 including all 20 zero-degree concepts)) at the registered reader with the corrected twin; prediction 3/4 held | stands; registered-reader scope; consequences executed on F-15 / F-17 / F-16 / F-18 / K-2 as pinned; DATED NOTE 2026-10-05 (F-28, STAGE LENS): through a balanced lens at S3 L3 the same cell reads ABSORBED with D +0.0019 [+0.0006, +0.0032] / +0.0004 [-0.0011, +0.0018], Mt +0.0042 [+0.0025, +0.0061] / +0.0021 [+0.0002, +0.0041] and Mn_r +0.0023 [+0.0010, +0.0038] / +0.0017 [-0.0000, +0.0035] (OFF3-ABSORBED, OFF4-ABSORBED) — the corrected twin point-positive on both harvests, CI-clear on the primary only at this lens, where it cost -0.0031 / -0.0067 at the lopsided one; the SPLIT at the lopsided lens stands as registered; DATED NOTE 2026-10-05 (F-28, STAGE READERS): the S3 L3 cell under the stronger readers — at the lopsided lens R-N (hz-rescaled) SURVIVES-STRONGER (EQUAL-DIMS-ONLY: Mt -0.0041 [-0.0063, -0.0018] / -0.0015 [-0.0055, +0.0029], Mn_r -0.0121 [-0.0141, -0.0100] / -0.0144 [-0.0173, -0.0116]; the fire is the twin's larger cost) (D +0.0080 [+0.0062, +0.0097] / +0.0130 [+0.0095, +0.0169]), R-T (two-tower) ABSORBED (D +0.0012 [-0.0003, +0.0026] / +0.0008 [-0.0006, +0.0023]; Mt +0.0011 [-0.0004, +0.0024] / -0.0019 [-0.0034, -0.0004]; Mn_r -0.0001 [-0.0015, +0.0014] / -0.0027 [-0.0045, -0.0009]); at the balanced lens R-N (hz-rescaled) NOT-SHOWN (D +0.0040 [+0.0025, +0.0054] / +0.0045 [+0.0027, +0.0061]; Mn_r -0.0046 [-0.0062, -0.0031] / -0.0071 [-0.0087, -0.0056]), R-T (two-tower) ABSORBED (D +0.0015 [+0.0003, +0.0027] / +0.0013 [-0.0002, +0.0028]); K-2 at S3 L3 stays OPEN at the registered reader (the SPLIT stands as registered); the K-2 word at S3 K-2-ISOLATED (from the R-N (hz-rescaled) cell at the lopsided L3; L1 itself ABSORBED under both readers); DATED NOTE 2026-10-09 (F-28, STAGE READERS-LONG): the S3 L3 cell under the longer-budget reader — at the lopsided lens R-L (longer budget) SURVIVES-STRONGER (EQUAL-DIMS-ONLY: Mt +0.0023 [+0.0008, +0.0038] / +0.0031 [+0.000336, +0.0058], Mn_r -0.0028 [-0.0043, -0.0011] / -0.0023 [-0.0045, -0.000098]; the fire is inflated by the twin's own cost) (D +0.005107 [+0.0034, +0.0069] / +0.005410 [+0.0034, +0.0078]); at the balanced lens R-L (longer budget) ABSORBED (D +0.0015 [+0.000220, +0.0029] / +0.0005 [-0.0007, +0.0018]; Mn_r +0.0010 [-0.000028, +0.0020] / +0.0012 [-0.000403, +0.0029]); K-2 at S3 L3 stays OPEN at the registered reader (the SPLIT stands as registered); the FINAL K-2 word at S3 over both reader stages K-2-ISOLATED |
    sideband/FINDINGS.md:62 · d147e8c0cd6b
  3. 2026-09-07sidebandF-15amendment

    [read under the 2026-09-07 rider amendment after the registered rider fired at S1 L3 (raw leg 2 +0.0132 against the 0.01 band); diagnostic branch AMENDABLE, mechanism TWIN-MATCHED; a post-hoc reroute around a fired kill condition: …

    The full line, as written
    - [read under the 2026-09-07 rider amendment after the registered rider fired at S1 L3 (raw leg 2 +0.0132 against the 0.01 band); diagnostic branch AMENDABLE, mechanism TWIN-MATCHED; a post-hoc reroute around a fired kill condition: the leg that fired was replaced after the fire by a twin-referenced dependence-excess leg whose S1 seed-0 values at both spaces were known from the diagnostic payload when the amendment was written; it is a fresh test only at the other three harvests; the E leg sees every non-causal block the raw leg sees (the twin lacks it) and is sharper under a drifted reader; what it drops (prediction-marginal drift and twin-shared position information) are non-leaks; S1 diagnostic X_raw at seed 0 +0.0006, the registered statistic, twin-referenced, would also have passed] DATED ERRATUM (2026-09-14; F-25, OQ-18): at L2 (offsets 2-3) the phrase "beyond the strongest free baseline" is corrected to "beyond the 657-summary PROXY of the free baseline; NOT beyond the full free baseline at the registered reader at the registered budget (OQ-18; every D CI95 upper below +0.005 on both harvests)" (F-25). Gate D +0.0011 [-0.0001, +0.0023] / +0.0006 [-0.0007, +0.0020] (per seed +0.0004 / -0.0002 / +0.0031 / +0.0003 / +0.0003 / +0.0012) with Mt +0.0015 [-0.0007, +0.0038] / +0.0022 [+0.0001, +0.0041] and Mn_r +0.0004 [-0.0019, +0.0025] / +0.0016 [-0.0004, +0.0035] beside it; linear off2 D +0.0016 [+0.0006, +0.0025] / +0.0022 [+0.0003, +0.0039]; horizons off3: OFF3-ABSORBED. The claim against the proxy stands as written; the L1 verdict (F-24) is untouched.
    sideband/FINDINGS.md:181 · d147e8c0cd6b
  4. 2026-09-07sidebandF-15amendment

    [read under the 2026-09-07 rider amendment after the registered rider fired at S1 L3 (raw leg 2 +0.0132 against the 0.01 band); diagnostic branch AMENDABLE, mechanism TWIN-MATCHED; a post-hoc reroute around a fired kill condition: …

    The full line, as written
    - [read under the 2026-09-07 rider amendment after the registered rider fired at S1 L3 (raw leg 2 +0.0132 against the 0.01 band); diagnostic branch AMENDABLE, mechanism TWIN-MATCHED; a post-hoc reroute around a fired kill condition: the leg that fired was replaced after the fire by a twin-referenced dependence-excess leg whose S1 seed-0 values at both spaces were known from the diagnostic payload when the amendment was written; it is a fresh test only at the other three harvests; the E leg sees every non-causal block the raw leg sees (the twin lacks it) and is sharper under a drifted reader; what it drops (prediction-marginal drift and twin-shared position information) are non-leaks; S1 diagnostic X_raw at seed 0 +0.0006, the registered statistic, twin-referenced, would also have passed] DATED ERRATUM (2026-09-14; F-25, OQ-18; wording only): at L3 (offsets 2-3) "beyond the strongest free baseline" becomes "beyond the 657-summary proxy; not re-demonstrated against the full free baseline at this power at L3 (OQ-18)" (F-25) at the L3 group lens (S1: one mega-group of 455/510 concepts including all 165 zero-degree concepts, the next group 10). Gate D +0.0010 [-0.0006, +0.0026] / +0.0012 [-0.0015, +0.0040] (per seed -0.0019 / -0.0015 / +0.0065 / +0.0012 / -0.0006 / +0.0031) with Mt +0.0013 [-0.0010, +0.0037] / +0.0028 [-0.0001, +0.0059] and Mn_r +0.0003 [-0.0019, +0.0026] / +0.0015 [-0.0013, +0.0044] beside it; linear off2 D -0.0028 [-0.0052, -0.0005] / +0.0029 [-0.0007, +0.0063]; horizons off3: OFF3-NOT-SHOWN. K-2 is OPEN at the full-free level at L3; no kill is claimed.
    sideband/FINDINGS.md:182 · d147e8c0cd6b
  5. 2026-09-07sidebandF-16amendment

    [read under the 2026-09-07 rider amendment after the registered rider fired at S1 L3 (raw leg 2 +0.0132 against the 0.01 band); diagnostic branch AMENDABLE, mechanism TWIN-MATCHED; a post-hoc reroute around a fired kill condition: …

    The full line, as written
    - [read under the 2026-09-07 rider amendment after the registered rider fired at S1 L3 (raw leg 2 +0.0132 against the 0.01 band); diagnostic branch AMENDABLE, mechanism TWIN-MATCHED; a post-hoc reroute around a fired kill condition: the leg that fired was replaced after the fire by a twin-referenced dependence-excess leg whose S1 seed-0 values at both spaces were known from the diagnostic payload when the amendment was written; it is a fresh test only at the other three harvests; the E leg sees every non-causal block the raw leg sees (the twin lacks it) and is sharper under a drifted reader; what it drops (prediction-marginal drift and twin-shared position information) are non-leaks; S1 diagnostic X_raw at seed 0 +0.0006, the registered statistic, twin-referenced, would also have passed] DATED NOTE (2026-09-14; F-25, OQ-18): the mechanism reading is moot for the S1 L2 margin at the registered reader (ABSORBED by the full free baseline; every D CI95 upper below +0.005 on both harvests); not rewritten.
    sideband/FINDINGS.md:191 · d147e8c0cd6b
  6. 2026-09-07sidebandF-17amendment

    [read under the 2026-09-07 rider amendment after the registered rider fired at S1 L3 (raw leg 2 +0.0132 against the 0.01 band); diagnostic branch AMENDABLE, mechanism TWIN-MATCHED; a post-hoc reroute around a fired kill condition: …

    The full line, as written
    - [read under the 2026-09-07 rider amendment after the registered rider fired at S1 L3 (raw leg 2 +0.0132 against the 0.01 band); diagnostic branch AMENDABLE, mechanism TWIN-MATCHED; a post-hoc reroute around a fired kill condition: the leg that fired was replaced after the fire by a twin-referenced dependence-excess leg whose S1 seed-0 values at both spaces were known from the diagnostic payload when the amendment was written; it is a fresh test only at the other three harvests; the E leg sees every non-causal block the raw leg sees (the twin lacks it) and is sharper under a drifted reader; what it drops (prediction-marginal drift and twin-shared position information) are non-leaks; S1 diagnostic X_raw at seed 0 +0.0006, the registered statistic, twin-referenced, would also have passed] DATED ERRATUM (2026-09-14; F-25, OQ-18): at L2 (offsets 2-3) the phrase "beyond the strongest free baseline" is corrected to "beyond the 657-summary PROXY of the free baseline; NOT beyond the full free baseline at the registered reader at the registered budget (OQ-18; every D CI95 upper below +0.005 on both harvests)" (F-25). Gate D +0.0018 [+0.0008, +0.0030] / +0.0003 [-0.0009, +0.0013] (per seed +0.0010 / +0.0024 / +0.0021 / -0.0001 / -0.0002 / +0.0011) with Mt +0.0021 [+0.0007, +0.0034] / +0.0018 [-0.0000, +0.0036] and Mn_r +0.0003 [-0.0011, +0.0015] / +0.0015 [-0.0002, +0.0031] beside it; linear off2 D +0.0010 [+0.0002, +0.0017] / +0.0002 [-0.0009, +0.0014]; horizons off3: OFF3-ABSORBED. acc(fplus, mlp256) - floor per seed +0.0329 / +0.0333 / +0.0340 (primary), +0.0265 / +0.0294 / +0.0305 (replication). The claim against the proxy stands as written; the L1 verdict (F-24) is untouched.
    sideband/FINDINGS.md:200 · d147e8c0cd6b
  7. 2026-09-07sidebandF-17amendment

    [read under the 2026-09-07 rider amendment after the registered rider fired at S1 L3 (raw leg 2 +0.0132 against the 0.01 band); diagnostic branch AMENDABLE, mechanism TWIN-MATCHED; a post-hoc reroute around a fired kill condition: …

    The full line, as written
    - [read under the 2026-09-07 rider amendment after the registered rider fired at S1 L3 (raw leg 2 +0.0132 against the 0.01 band); diagnostic branch AMENDABLE, mechanism TWIN-MATCHED; a post-hoc reroute around a fired kill condition: the leg that fired was replaced after the fire by a twin-referenced dependence-excess leg whose S1 seed-0 values at both spaces were known from the diagnostic payload when the amendment was written; it is a fresh test only at the other three harvests; the E leg sees every non-causal block the raw leg sees (the twin lacks it) and is sharper under a drifted reader; what it drops (prediction-marginal drift and twin-shared position information) are non-leaks; S1 diagnostic X_raw at seed 0 +0.0006, the registered statistic, twin-referenced, would also have passed] DATED STATUS (2026-09-14; F-25, OQ-18): at L3 (offsets 2-4) the claim against the proxy stands; against the FULL free baseline it is OPEN — SPLIT, the gate fired on the replication harvest only at the L3 group lens (S3: two mega-groups of 268 and 193 = 461/510 including all 20 zero-degree concepts): Gate D +0.0041 [+0.0026, +0.0055] / +0.0053 [+0.0019, +0.0091] (per seed +0.0025 / +0.0033 / +0.0064 / +0.0029 / +0.0071 / +0.0059) with Mt +0.0010 [-0.0012, +0.0030] / -0.0014 [-0.0041, +0.0016] and Mn_r -0.0031 [-0.0053, -0.0010] / -0.0067 [-0.0090, -0.0042] beside it; linear off2 D +0.0073 [+0.0050, +0.0095] / +0.0049 [+0.0022, +0.0078]; horizons off3: OFF3-SURVIVES, off4: OFF4-NOT-SHOWN. No relabeling either way; a third harvest is the only resolution and is not authorized.
    sideband/FINDINGS.md:201 · d147e8c0cd6b
  8. 2026-09-07sidebandF-18amendment

    [read under the 2026-09-07 rider amendment after the registered rider fired at S1 L3 (raw leg 2 +0.0132 against the 0.01 band); diagnostic branch AMENDABLE, mechanism TWIN-MATCHED; a post-hoc reroute around a fired kill condition: …

    The full line, as written
    - [read under the 2026-09-07 rider amendment after the registered rider fired at S1 L3 (raw leg 2 +0.0132 against the 0.01 band); diagnostic branch AMENDABLE, mechanism TWIN-MATCHED; a post-hoc reroute around a fired kill condition: the leg that fired was replaced after the fire by a twin-referenced dependence-excess leg whose S1 seed-0 values at both spaces were known from the diagnostic payload when the amendment was written; it is a fresh test only at the other three harvests; the E leg sees every non-causal block the raw leg sees (the twin lacks it) and is sharper under a drifted reader; what it drops (prediction-marginal drift and twin-shared position information) are non-leaks; S1 diagnostic X_raw at seed 0 +0.0006, the registered statistic, twin-referenced, would also have passed] DATED NOTE (2026-09-14; F-25, OQ-18): the mechanism reading is moot for the S3 L2 margin at the registered reader (ABSORBED by the full free baseline; every D CI95 upper below +0.005 on both harvests); not rewritten.
    sideband/FINDINGS.md:210 · d147e8c0cd6b
  9. 2026-09-07sidebandF-25amendment

    amendment: [read under the 2026-09-07 rider amendment after the registered rider fired at S1 L3 (raw leg 2 +0.0132 against the 0.01 band); diagnostic branch AMENDABLE, mechanism TWIN-MATCHED; a post-hoc reroute around a fired kill condition: …

    The full line, as written
    - amendment: [read under the 2026-09-07 rider amendment after the registered rider fired at S1 L3 (raw leg 2 +0.0132 against the 0.01 band); diagnostic branch AMENDABLE, mechanism TWIN-MATCHED; a post-hoc reroute around a fired kill condition: the leg that fired was replaced after the fire by a twin-referenced dependence-excess leg whose S1 seed-0 values at both spaces were known from the diagnostic payload when the amendment was written; it is a fresh test only at the other three harvests; the E leg sees every non-causal block the raw leg sees (the twin lacks it) and is sharper under a drifted reader; what it drops (prediction-marginal drift and twin-shared position information) are non-leaks; S1 diagnostic X_raw at seed 0 +0.0006, the registered statistic, twin-referenced, would also have passed] RIDER REPORT under the amendment (leg 1 passed and the amended leg 2 passed at all eight harvest x space cells; the REGISTERED raw leg 2 is REPORTED, never halting): S1 L2 raw +0.0024, amended dE_br +0.0005 [-0.0001, +0.0011]; S1 L3 raw +0.0132 (ABOVE the band), amended dE_br +0.0026 [-0.0004, +0.0054]; S1_r L2 raw +0.0105 (ABOVE the band), amended dE_br -0.0005 [-0.0011, +0.0002]; S1_r L3 raw -0.0024, amended dE_br -0.0017 [-0.0046, +0.0012]; S3 L2 raw -0.0001, amended dE_br +0.0001 [-0.0000, +0.0002]; S3 L3 raw +0.0020, amended dE_br +0.0007 [-0.0021, +0.0038]; S3_r L2 raw +0.0014, amended dE_br +0.0001 [-0.0000, +0.0001]; S3_r L3 raw -0.0007, amended dE_br +0.0007 [-0.0023, +0.0040]. At S1 L3, the cell the reroute was written around, the amended pass is NOT a CI-clear null: dE_br +0.0026 has upper +0.0054, at or above the +0.005 materiality floor, the only cell of eight where that holds; it passes because its lower bound -0.0004 is below zero. The diagnostic's pooled seeds 0-2 at that cell read dE_br +0.0018 [-0.0004, +0.0039] (upper below the floor, lower below zero): a power statement, not an exclusion. The registered raw leg is above its 0.01 band at TWO cells: S1 L3 (+0.0132, the fire that stopped the registered run) and S1_r L2 (+0.0105). COUNTERFACTUAL, stated as one: the frozen instrument was never run on the S1 replication; its grid trains a space's shuffled cells first and its rider raises SystemExit on a leg-2 fail before that space's first unshuffled cell, and the retrained shuffled rows in the committed state files reproduce the frozen runs' accuracies exactly (all 18 retrained rows across S1, S3 and S3_r equal their frozen rows' test_acc, max abs diff 0.0; S1 L3 seed 0 dE_br +0.0026 = the diagnostic's), so under the frozen rider the S1 replication WOULD HAVE halted at L2 and OQ-18 would have had NO S1 label. The amended leg at S1_r L2 reads dE_br -0.0005 [-0.0011, +0.0002] with the aligned arm's own dependence excess over fplus -0.0009: raw up, dependence down — prediction-marginal drift, one of the two non-leak mechanisms the amendment named (the S1 L3 fire itself read TWIN-MATCHED); this is the one place the reroute is exercised as a FRESH test, and the registered statistic fired there on drift alone. DIRECTION OF THE REROUTE, for the record: the registered instrument would have read no S1 label; the amendment enabled reading a kill-extension (S1 L2 ABSORBED, K-2 FIRED at L2) and a NOT-SHOWN, not a survival. RUNS: S1 primary resumed from the fired run's committed state file (resumed_cells 30, retrained_shuffled_rows 6); S1 replication fresh (resumed_cells 0); S3 primary and replication resumed from their frozen-instrument runs (resumed_cells 69 / 69, retrained_shuffled_rows 6 / 6) with the frozen-instrument payloads preserved, NEVER concluded, as results/real/fullfree_coarse_s3.rider_v1.json (sha 8fc04f277c72...) and fullfree_coarse_s3_r.rider_v1.json (sha 33e73c080f94...). ATTESTATION: all four payloads record HEAD 3ec1064-dirty over exactly the one licensed tracked modification (the S1 state file's own append), git_status_licensed true; amended instrument sha aafd3ae84998... importing the frozen instrument byte-identical (sha 5efa7063767b...); labels read only through the amended --conclude-from, which refuses v1 payloads and unlicensed attestations. PROCESS NOTE (disclosed, not omitted): before the S1 conclude ran, a key-scrubbed print of the S1 primary payload's rider block overran into the S1 L2 gate row (its fire flag and D margins were seen); no label string and no v1 payload were read; the labels, the conclude step and the pinned texts are mechanical and unaffected.
    sideband/FINDINGS.md:282 · d147e8c0cd6b
  10. 2026-09-07sidebandF-25dated note

    source: results/real/fullfree_coarse_s1.json, fullfree_coarse_s1.json.state.jsonl, fullfree_coarse_s1_r.json, fullfree_coarse_s1_r.json.state.jsonl, fullfree_coarse_s3.json, fullfree_coarse_s3.json.state.jsonl, fullfree_coarse_s3_r.json, fullfree_coarse_s3_r.json.state.jsonl (sha-header bound to their harvests; …

    The full line, as written
    - source: results/real/fullfree_coarse_s1.json, fullfree_coarse_s1.json.state.jsonl, fullfree_coarse_s1_r.json, fullfree_coarse_s1_r.json.state.jsonl, fullfree_coarse_s3.json, fullfree_coarse_s3.json.state.jsonl, fullfree_coarse_s3_r.json, fullfree_coarse_s3_r.json.state.jsonl (sha-header bound to their harvests; committed with this finding); tools/run_fullfree_coarse_amended.py (the amended instrument, sha aafd3ae84998..., importing tools/run_fullfree_coarse.py byte-identical at its frozen sha) and tests/test_fullfree_coarse_amended.py (the committed amended payloads reproduce their own conclusion, resume plan, attestation and v1 preservation; the frozen test file skips amended payloads by a dated note); results/real/fullfree_coarse_s3.rider_v1.json, fullfree_coarse_s3_r.rider_v1.json (frozen-instrument payloads, preserved, never concluded); results/real/oq18_rider_diag_s1.json + state file and docs/OQ18_RIDER_DIAGNOSTIC.md (the fire's diagnostic); results/real/oq18_rider_poscontrol_s1.json + state file and docs/OQ18_RIDER_POSCONTROL.md (the positive control); docs/BLUEPRINT.md dated amendment bullet (2026-09-07 / 2026-09-13); docs/BLUEPRINT.md OQ-18 registration and OUTCOME blocks; the committed Phase-3 state files for the paired linear cells (results/real/probes_s1{,_replication}.json.state.jsonl, probes_s3{,_replication}.json.state.jsonl).
    sideband/FINDINGS.md:285 · d147e8c0cd6b
  11. 2026-09-07sidebandF-25amendment

    finding: [read under the 2026-09-07 rider amendment after the registered rider fired at S1 L3 (raw leg 2 +0.0132 against the 0.01 band); diagnostic branch AMENDABLE, mechanism TWIN-MATCHED; a post-hoc reroute around a fired kill condition: …

    The full line, as written
    - finding: [read under the 2026-09-07 rider amendment after the registered rider fired at S1 L3 (raw leg 2 +0.0132 against the 0.01 band); diagnostic branch AMENDABLE, mechanism TWIN-MATCHED; a post-hoc reroute around a fired kill condition: the leg that fired was replaced after the fire by a twin-referenced dependence-excess leg whose S1 seed-0 values at both spaces were known from the diagnostic payload when the amendment was written; it is a fresh test only at the other three harvests; the E leg sees every non-causal block the raw leg sees (the twin lacks it) and is sharper under a drifted reader; what it drops (prediction-marginal drift and twin-shared position information) are non-leaks; S1 diagnostic X_raw at seed 0 +0.0006, the registered statistic, twin-referenced, would also have passed] MECHANICAL LABELS per (source, space), both harvests, registered reader at the registered budget: S1 L2 ABSORBED; S1 L3 NOT-SHOWN; S3 L2 ABSORBED; S3 L3 SPLIT (every L3 label read at the L3 group lens (S1: one mega-group of 455/510 concepts including all 165 zero-degree concepts, the next group 10; S3: two mega-groups of 268 and 193 = 461/510 including all 20 zero-degree concepts)). S1 L2: ABSORBED. Gate D +0.0011 [-0.0001, +0.0023] / +0.0006 [-0.0007, +0.0020] (per seed +0.0004 / -0.0002 / +0.0031 / +0.0003 / +0.0003 / +0.0012) with Mt +0.0015 [-0.0007, +0.0038] / +0.0022 [+0.0001, +0.0041] and Mn_r +0.0004 [-0.0019, +0.0025] / +0.0016 [-0.0004, +0.0035] beside it; linear off2 D +0.0016 [+0.0006, +0.0025] / +0.0022 [+0.0003, +0.0039]; horizons off3: OFF3-ABSORBED. DATED NOTE 2026-10-16 (image audit 2026-10-15, section B item 2; docs/audits/2026-10-15-image-audit/seed_var.py rebuilds every interval of this entry from the committed state files to 1e-12 and brackets each with three seed-inclusive approximations; NON-CRITERION, no label moves): every interval here is a prompt-cluster bootstrap over counts summed across the three train seeds, so it is conditional on the three trained probes per arm, and the registered all-seeds rule protects the fire direction only. The deciding row of S1 L2 ABSORBED, the replication's linear off2 D +0.0022 [+0.0003, +0.0039] (per seed +0.0050 / +0.0009 / +0.0006), reads [-0.0012, +0.0055] under the finder's normal combination of the bootstrap and seed SEs, [-0.00065, +0.00498] under an interaction-corrected method of moments (that upper stays just below the floor) and [-0.0002, +0.0064] under a prompt x seed pigeonhole bootstrap: its upper reaches the +0.005 floor under two of the three, so S1 L2 ABSORBED is seed-sensitive at that row; the S3 L2 D uppers stay below the floor under all three (largest +0.0032 / +0.0030 / +0.0035) and the S3 L3 replication gate cell's D lower stays above zero under all three (+0.0009 / +0.0017 / +0.0006; SPLIT stands); the primary S3 L3 Mn_r -0.0031 [-0.0053, -0.0010], on which the post-hoc "same shape" sentence below rests, reads [-0.0075, +0.0014] / [-0.0071, +0.0010] / [-0.0072, +0.0010], its upper above zero under all three (the replication's stays below zero), and the off3 and off4 Mt rows were not bracketed. Three seeds give two degrees of freedom: the component is bracketed, not measured; measuring it needs a newly registered multi-seed run. S1 L3: NOT-SHOWN at the L3 group lens (S1: one mega-group of 455/510 concepts including all 165 zero-degree concepts, the next group 10). Gate D +0.0010 [-0.0006, +0.0026] / +0.0012 [-0.0015, +0.0040] (per seed -0.0019 / -0.0015 / +0.0065 / +0.0012 / -0.0006 / +0.0031) with Mt +0.0013 [-0.0010, +0.0037] / +0.0028 [-0.0001, +0.0059] and Mn_r +0.0003 [-0.0019, +0.0026] / +0.0015 [-0.0013, +0.0044] beside it; linear off2 D -0.0028 [-0.0052, -0.0005] / +0.0029 [-0.0007, +0.0063]; horizons off3: OFF3-NOT-SHOWN. S3 L2: ABSORBED. Gate D +0.0018 [+0.0008, +0.0030] / +0.0003 [-0.0009, +0.0013] (per seed +0.0010 / +0.0024 / +0.0021 / -0.0001 / -0.0002 / +0.0011) with Mt +0.0021 [+0.0007, +0.0034] / +0.0018 [-0.0000, +0.0036] and Mn_r +0.0003 [-0.0011, +0.0015] / +0.0015 [-0.0002, +0.0031] beside it; linear off2 D +0.0010 [+0.0002, +0.0017] / +0.0002 [-0.0009, +0.0014]; horizons off3: OFF3-ABSORBED. acc(fplus, mlp256) - floor per seed +0.0329 / +0.0333 / +0.0340 (primary), +0.0265 / +0.0294 / +0.0305 (replication). S3 L3: SPLIT at the L3 group lens (S3: two mega-groups of 268 and 193 = 461/510 including all 20 zero-degree concepts). Gate D +0.0041 [+0.0026, +0.0055] / +0.0053 [+0.0019, +0.0091] (per seed +0.0025 / +0.0033 / +0.0064 / +0.0029 / +0.0071 / +0.0059) with Mt +0.0010 [-0.0012, +0.0030] / -0.0014 [-0.0041, +0.0016] and Mn_r -0.0031 [-0.0053, -0.0010] / -0.0067 [-0.0090, -0.0042] beside it; linear off2 D +0.0073 [+0.0050, +0.0095] / +0.0049 [+0.0022, +0.0078]; horizons off3: OFF3-SURVIVES, off4: OFF4-NOT-SHOWN. PREDICTION CHECK (pinned before the smokes): S1 L2 predicted ABSORBED, read ABSORBED (held); S1 L3 predicted NOT-SHOWN, read NOT-SHOWN (held); S3 L2 predicted ABSORBED, read ABSORBED (held); S3 L3 predicted SURVIVES-FULL-FREE, read SPLIT (FAILED) — 3 of 4 held. SUB-CLAUSES of the pinned prediction, scored: S1 L2 gate D within +/-0.003 of zero HELD (+0.0011 / +0.0006); every D upper below the floor HELD; Mt +0.002..+0.006 FAILED on the primary (+0.0015) and HELD on the replication (+0.0022). S1 L3 gate D point below +0.005 with at least one D upper at or above it HELD. S3 L3 SURVIVES at off2 with D +0.005..+0.020 on both harvests and Mt positive FAILED (read SPLIT: primary D +0.0041, replication Mt -0.0014); OFF3-NOT-SHOWN or OFF3-SURVIVES HELD (OFF3-SURVIVES); OFF4-ABSORBED FAILED (OFF4-NOT-SHOWN: replication off4 D +0.0035 [+0.0009, +0.0064]). Mf CI-clear positive at every gate cell HELD (lowers S1 L2 +0.0371, S1 L3 +0.0804, S1_r L2 +0.0350, S1_r L3 +0.0724, S3 L2 +0.0268, S3 L3 +0.0830, S3_r L2 +0.0219, S3_r L3 +0.0765). S3 L3 SPLIT was named in advance among the outcomes that would surprise. PROVENANCE: fullfree_coarse_s1.json: git 3ec10645a8e1, 60 trained cells, resumed 30, hz moments == committed OQ-17 True; fullfree_coarse_s1_r.json: git 3ec10645a8e1, 60 trained cells, resumed 0, hz moments == committed OQ-17 True; fullfree_coarse_s3.json: git 3ec10645a8e1, 69 trained cells, resumed 69, hz moments == committed OQ-17 True; fullfree_coarse_s3_r.json: git 3ec10645a8e1, 69 trained cells, resumed 69, hz moments == committed OQ-17 True. PRECONDITIONS AND RIDERS: fullfree_coarse_s1.json L2: rider PASS, reader valid, min gate acc - floor +0.0115, Mf upper +0.0557, Mf_x +0.0477, best epochs {'fplus': [0, 0, 1], 'fplus_b': [0, 1, 1], 'fplus_r': [0, 1, 1], 'a': [9, 12, 22]}; fullfree_coarse_s1.json L3: rider PASS, reader valid, min gate acc - floor +0.0643, Mf upper +0.0946, Mf_x +0.0961, best epochs {'fplus': [2, 2, 4], 'fplus_b': [2, 2, 2], 'fplus_r': [0, 2, 2], 'a': [15, 29, 29]}; fullfree_coarse_s1_r.json L2: rider PASS, reader valid, min gate acc - floor +0.0074, Mf upper +0.0527, Mf_x +0.0433, best epochs {'fplus': [0, 1, 1], 'fplus_b': [1, 1, 1], 'fplus_r': [1, 1, 1], 'a': [9, 12, 14]}; fullfree_coarse_s1_r.json L3: rider PASS, reader valid, min gate acc - floor +0.0646, Mf upper +0.0931, Mf_x +0.0912, best epochs {'fplus': [1, 2, 2], 'fplus_b': [1, 1, 2], 'fplus_r': [1, 2, 2], 'a': [16, 25, 29]}; fullfree_coarse_s3.json L2: rider PASS, reader valid, min gate acc - floor +0.0015, Mf upper +0.0366, Mf_x +0.0311, best epochs {'fplus': [1, 1, 1], 'fplus_b': [1, 1, 1], 'fplus_r': [1, 1, 1], 'a': [14, 19, 19]}; fullfree_coarse_s3.json L3: rider PASS, reader valid, min gate acc - floor +0.1087, Mf upper +0.1005, Mf_x +0.1153, best epochs {'fplus': [1, 1, 1], 'fplus_b': [1, 1, 1], 'fplus_r': [1, 1, 1], 'a': [19, 22, 25]}; fullfree_coarse_s3_r.json L2: rider PASS, reader valid, min gate acc - floor +0.0016, Mf upper +0.0325, Mf_x +0.0270, best epochs {'fplus': [1, 1, 1], 'fplus_b': [1, 1, 2], 'fplus_r': [1, 1, 2], 'a': [6, 7, 11]}; fullfree_coarse_s3_r.json L3: rider PASS, reader valid, min gate acc - floor +0.1042, Mf upper +0.0956, Mf_x +0.1038, best epochs {'fplus': [1, 1, 1], 'fplus_b': [1, 2, 2], 'fplus_r': [1, 2, 2], 'a': [21, 23, 29]}. CONSEQUENCES EXECUTED per the pinned rules, dated, never rewrites: the per-space bullets on F-15 / F-17, the K-2 STATUS block, the F-16 / F-18 notes where ABSORBED, the scorecard cells, the REPORT / README notes, the OQ-18 OUTCOME texts.
    sideband/FINDINGS.md:286 · d147e8c0cd6b
  12. 2026-09-06sidebandfindings table row F-24dated note

    | F-24 | 2026-09-06 | OQ-17 FULL-FREE EXISTENCE RE-TEST: F-15 and F-17 margins ABSORBED by the full free baseline ((a) ++ hashed token ids ++ z-scored h_a) at the registered reader at the pinned L1 offset-2 cells of both sources — …

    The full line, as written
    | F-24 | 2026-09-06 | OQ-17 FULL-FREE EXISTENCE RE-TEST: F-15 and F-17 margins ABSORBED by the full free baseline ((a) ++ hashed token ids ++ z-scored h_a) at the registered reader at the pinned L1 offset-2 cells of both sources — gate D S1 +0.0002 [-0.0012, +0.0015] / +0.0012 [-0.0003, +0.0028], S3 +0.0003 [-0.0013, +0.0020] / +0.0001 [-0.0019, +0.0021], every D upper below +0.005, Mt about +0.003 beside it; prediction failed on both sources; corrected twin TWIN-ROBUST (S1) / TWIN-ROBUST (S3) | stands (criterion-bearing for F-15 / F-17 through the pinned OQ-17 consequence rules; critic-recomputed three times); DATED NOTE 2026-10-05 (F-28, STAGE READERS): the S3 L1 kill re-tested under the hz-rescaled and two-tower readers — R-N (hz-rescaled) ABSORBED (D +0.0030 [+0.0016, +0.0045] / +0.0032 [+0.0016, +0.004951]; Mt -0.0016 [-0.0033, +0.0001] / -0.0005 [-0.0022, +0.0014]; Mn_r -0.0046 [-0.0063, -0.0028] / -0.0037 [-0.0052, -0.0020]), R-T (two-tower) ABSORBED (D +0.0006 [-0.0002, +0.0014] / +0.0007 [-0.0002, +0.0015]; Mt +0.0026 [+0.0009, +0.0041] / +0.0009 [-0.0006, +0.0024]; Mn_r +0.0020 [+0.0005, +0.0035] / +0.0002 [-0.0009, +0.0014]); the K-2 word at S3 K-2-ISOLATED (from the R-N (hz-rescaled) cell at the lopsided L3; L1 itself ABSORBED under both readers); the registered-reader kill stands as a statement about that reader; DATED NOTE 2026-10-09 (F-28, STAGE READERS-LONG): the S3 L1 kill re-tested under the longer-budget reader — R-L (longer budget) ABSORBED (D +0.0012 [+0.000250, +0.0022] / +0.0015 [-0.000044, +0.0032]; Mt +0.0017 [+0.000095, +0.0032] / +0.0023 [+0.0010, +0.0038]; Mn_r +0.000487 [-0.0011, +0.0020] / +0.0008 [-0.0005, +0.0021]); the FINAL K-2 word at S3 over both reader stages K-2-ISOLATED (from the R-L (longer budget) and R-N (hz-rescaled) cells at the lopsided L3; L1 itself ABSORBED under all three readers); the registered-reader kill stands as a statement about that reader; DATED NOTE 2026-10-17 (F-29, S1 READER RE-TEST): the S1 gate cell re-tested under the hz-rescaled, two-tower and longer-budget readers — R-N (hz-rescaled) ABSORBED (D +0.0026 [+0.0008, +0.0043] / +0.0022 [+0.000497, +0.0039]); R-T (two-tower) ABSORBED (D +0.000339 [-0.000377, +0.0011] / +0.0011 [+0.000363, +0.0019]); R-L (longer budget) ABSORBED (D +0.000382 [-0.0009, +0.0016] / -0.000044 [-0.0014, +0.0013]); the S1 word K-2-NOT-OVERTURNED-AT-S1 (under readers R-N, R-T and R-L at S1; one more family of readers, not K-2's universal); F-24's own S1 words stand as the registered reader's |
    sideband/FINDINGS.md:61 · d147e8c0cd6b
  13. 2026-09-06sidebandF-15erratum

    DATED ERRATUM (2026-09-06; F-24, OQ-17): the phrase "beyond the strongest free baseline" in this finding's title and text is corrected to "beyond the 657-summary PROXY of the free baseline; NOT beyond the full free baseline at the registered reader at the registered budget at the pinned cells L1 offsets 2-3 (OQ-17, F-24; …

    The full line, as written
    - DATED ERRATUM (2026-09-06; F-24, OQ-17): the phrase "beyond the strongest free baseline" in this finding's title and text is corrected to "beyond the 657-summary PROXY of the free baseline; NOT beyond the full free baseline at the registered reader at the registered budget at the pinned cells L1 offsets 2-3 (OQ-17, F-24; every D CI95 upper below +0.005 on both harvests)" — L2 / L3 and the further offsets stand against the proxy and are untested against the full free baseline; the claim against the proxy stands as written. Gate D +0.0002 [-0.0012, +0.0015] / +0.0012 [-0.0003, +0.0028] with Mt +0.0034 [+0.0021, +0.0048] / +0.0033 [+0.0012, +0.0052] beside it. The registered equal-dims twin was found NON-INERT after the fact (it reads the same row 17 positions ahead, beyond the label position for about 94% of anchors; Mn +0.0033 [+0.0018, +0.0049] on the primary harvest, CI-clear, and +0.0020 [-0.0000, +0.0041] on the replication, straddling), so the label carries the corrected-twin reading: TWIN-ROBUST — D_r +0.0002 [-0.0009, +0.0014] primary (seeds +0.0013 / -0.0016 / +0.0008), +0.0013 [-0.0002, +0.0028] replication (seeds +0.0029 / +0.0009 / -0.0001); Mn_r +0.0032 [+0.0016, +0.0050] / +0.0020 [+0.0002, +0.0037]; T +0.0000 [-0.0012, +0.0013] / +0.0000 [-0.0016, +0.0019] (TWIN-INFORMATION-NOT-SHOWN). On the primary and replication the corrected twin itself gains CI-clear over the base while T's CI includes zero, so the label-consumption mechanism is WITHDRAWN for both harvests: the shift twin's gain is matched by a prompt-independent block, which under the pinned reading means a same-marginal sparse block of these dims yields about Mt at this reader — the strongest form of the null, under which Mt stops being evidence of telemetry information (the registered reading of one construction, not a measurement of every block); D_r remains the matched comparison. The consequences act exactly as pinned; the null reproduces at a prompt-independent twin at the gate cell. The corrected twin ran the offset-2 mlp256 gate cell only, so the offset-3 leg of this scope and the linear rows rest on the registered shift twin.
    sideband/FINDINGS.md:180 · d147e8c0cd6b
  14. 2026-09-06sidebandF-16dated note

    DATED NOTE (2026-09-06; F-24): the shift-17 equal-dims twin used here as the noise control is NOT inert under a null direction — it reads a position whose forward pass has consumed the label token for about 94% of anchors and gains CI-clear over the base at the OQ-17 gate on the primary harvest (Mn +0.0033 [+0.0018, +0.0049]; …

    The full line, as written
    - DATED NOTE (2026-09-06; F-24): the shift-17 equal-dims twin used here as the noise control is NOT inert under a null direction — it reads a position whose forward pass has consumed the label token for about 94% of anchors and gains CI-clear over the base at the OQ-17 gate on the primary harvest (Mn +0.0033 [+0.0018, +0.0049]; replication +0.0020 [-0.0000, +0.0041], straddling); the "could only make M2n an UNDERestimate" caveat holds under SURVIVES only. The mechanism reading of this finding is moot for the L1 offsets 2-3 margin at the registered reader (a margin below the materiality floor and not beyond the equal-dims twin has no mechanism to price); the histogram-level rejection of topic-state as an explanation of the PROXY margin is untouched. Not rewritten.
    sideband/FINDINGS.md:190 · d147e8c0cd6b
  15. 2026-09-06sidebandF-17erratum

    DATED ERRATUM (2026-09-06; F-24, OQ-17): the phrase "beyond the strongest free baseline" in this finding's title and text is corrected to "beyond the 657-summary PROXY of the free baseline; NOT beyond the full free baseline at the registered reader at the registered budget at the pinned cells L1 offset 2 (OQ-17, F-24; …

    The full line, as written
    - DATED ERRATUM (2026-09-06; F-24, OQ-17): the phrase "beyond the strongest free baseline" in this finding's title and text is corrected to "beyond the 657-summary PROXY of the free baseline; NOT beyond the full free baseline at the registered reader at the registered budget at the pinned cells L1 offset 2 (OQ-17, F-24; every D CI95 upper below +0.005 on both harvests)" — L2 / L3 and the further offsets stand against the proxy and are untested against the full free baseline; the claim against the proxy stands as written. Gate D +0.0003 [-0.0013, +0.0020] / +0.0001 [-0.0019, +0.0021] with Mt +0.0033 [+0.0013, +0.0054] / +0.0013 [-0.0008, +0.0035] beside it. The registered equal-dims twin was found NON-INERT after the fact (it reads the same row 17 positions ahead, beyond the label position for about 94% of anchors; Mn +0.0030 [+0.0014, +0.0046] on the primary harvest, CI-clear, and +0.0012 [-0.0004, +0.0026] on the replication, straddling), so the label carries the corrected-twin reading: TWIN-ROBUST — D_r +0.0000 [-0.0014, +0.0014] primary (seeds +0.0010 / +0.0001 / -0.0009), +0.0008 [-0.0014, +0.0029] replication (seeds +0.0019 / +0.0009 / -0.0006); Mn_r +0.0033 [+0.0015, +0.0050] / +0.0005 [-0.0012, +0.0022]; T -0.0003 [-0.0021, +0.0015] / +0.0007 [-0.0005, +0.0018] (TWIN-INFORMATION-NOT-SHOWN). On the primary the corrected twin itself gains CI-clear over the base while T's CI includes zero, so the label-consumption mechanism is WITHDRAWN for that harvest: the shift twin's gain is matched by a prompt-independent block, which under the pinned reading means a same-marginal sparse block of these dims yields about Mt at this reader — the strongest form of the null, under which Mt stops being evidence of telemetry information (the registered reading of one construction, not a measurement of every block); D_r remains the matched comparison. On the replication the mechanism is neither withdrawn nor supported (Mn_r and T straddle zero). The consequences act exactly as pinned; the null reproduces at a prompt-independent twin at the gate cell. The corrected twin ran the mlp256 gate cell only; the linear row keeps its CI-positive-immaterial qualifier.
    sideband/FINDINGS.md:199 · d147e8c0cd6b
  16. 2026-09-06sidebandF-18dated note

    DATED NOTE (2026-09-06; F-24): the shift-17 equal-dims twin used here as the noise control is NOT inert under a null direction — it reads a position whose forward pass has consumed the label token for about 94% of anchors and gains CI-clear over the base at the OQ-17 gate on the primary harvest (Mn +0.0030 [+0.0014, +0.0046]; …

    The full line, as written
    - DATED NOTE (2026-09-06; F-24): the shift-17 equal-dims twin used here as the noise control is NOT inert under a null direction — it reads a position whose forward pass has consumed the label token for about 94% of anchors and gains CI-clear over the base at the OQ-17 gate on the primary harvest (Mn +0.0030 [+0.0014, +0.0046]; replication +0.0012 [-0.0004, +0.0026], straddling); the "could only make M2n an UNDERestimate" caveat holds under SURVIVES only. The mechanism reading of this finding is moot for the L1 offset 2 margin at the registered reader (a margin below the materiality floor and not beyond the equal-dims twin has no mechanism to price); the histogram-level rejection of topic-state as an explanation of the PROXY margin is untouched. Not rewritten.
    sideband/FINDINGS.md:209 · d147e8c0cd6b
  17. 2026-09-05sidebandfindings table row F-23erratum

    | F-23 | 2026-09-05 | OQ-16 CLOSED BY DATED ERRATUM (TARGET-FREE): the anchor state h_a is an exact linear function of the anchor's full next-token distribution, free under D-5; Stage-A telemetry arms struck pre-GPU; no partition branch claimed | stands (terminal for OQ-16; …

    The full line, as written
    | F-23 | 2026-09-05 | OQ-16 CLOSED BY DATED ERRATUM (TARGET-FREE): the anchor state h_a is an exact linear function of the anchor's full next-token distribution, free under D-5; Stage-A telemetry arms struck pre-GPU; no partition branch claimed | stands (terminal for OQ-16; critic-recomputed) |
    sideband/FINDINGS.md:60 · d147e8c0cd6b
  18. 2026-09-05sidebandF-23other

    regime: MATHEMATICAL IDENTITY, numerically checked on REAL byproduct rows (Llama-3.1-8B-Instruct lm_head, untied and bias-free, read from the cached bf16 checkpoint as float32; …

    The full line, as written
    - regime: MATHEMATICAL IDENTITY, numerically checked on REAL byproduct rows (Llama-3.1-8B-Instruct lm_head, untied and bias-free, read from the cached bf16 checkpoint as float32; 48 pinned rows = the first 4 blocks of the first 3 prompts of each of the four registered vanilla gamma-4 verdict runs, every npz sha-bound to its committed summary; CPU compute on the GPU box — the tool needs the cached weights and the gitignored raws; CI imports the module through tests/test_ha_identity.py, where torch and safetensors load lazily, and never runs it; the identity is generic to any linear unembedding of full column rank with vocab > d, tied or untied, and is measured here on one model). CLAIM: h_a (the final-norm output at the anchor, the registered Stage-R target) is determined by the anchor's logits, and therefore by the anchor's full next-token distribution p — the quantity D-5 names as the mandatory free baseline ("steering from the target's distribution at the last verified position, which every speculative decoder computes anyway"). W_U @ h reproduces the logged argmax 48/48 and the logged top-32 ids at 0.9948 mean overlap (min 0.969; max top-1 probability gap 0.0206 against the values logged by the live bf16 lm_head of the NF4-loaded verifier); the least-squares inverse from the float32 logits recovers h_a at median relative error 2.5e-8 (max 3.2e-8), the float32 rounding floor (from float64 logits 1.9e-14); closed-form least-squares recovery from only the top-k (id, log-prob) pairs of the recomputed logits (not the logged float16 pairs) reaches median per-row R^2 +1.0000 at k = 4097, +0.506 at 2048, +0.216 at 512, +0.078 at 128, +0.014 at 32 (median cosines 1.000 / 0.711 / 0.465 / 0.280 / 0.123) — a per-row R^2 about the row's own mean in raw h space from a prior-free solve, NOT Stage-R's standardized pooled-MSE R^2 and not comparable to the ridge arms' +0.258 / +0.324; the top-k recovery uses log-probabilities only with the log-partition as an extra unknown, so p alone, not the logits, determines h_a. CONDITIONING: W_U has full column rank (singular values 1.453 .. 94.5, condition 65); the all-ones vector, the softmax's free direction, lies 99.2% inside col(W_U), yet the augmented system [W_U, -1] has smallest singular value 1.451, so nothing is weakly determined; from bfloat16-rounded full log-probabilities with logZ unknown, h_a recovers to median relative error 3.5e-03 (max 6.9e-03). WHAT STANDS: every Stage-R and DR-1..DR-4 readout as pinned (F-22) — the gates were run against their registered arms and recomputed bit-exact. WHAT FALLS: their meaning for the lane. The registered FREE arm (657 anchor summaries; R^2 +0.083) and the DR-2/DR-4 token encodings (FREETOK +0.258, FREETOK2 +0.255) were lossy proxies for the free information D-5 names, from which the target follows exactly (R^2 1.000); the stream's reconstruction (TEL +0.253; UNION +0.324) is therefore K-2's "lossy copy of information the decoder already has", with no hashing caveat and no further comparator needed — "beyond p" is priced at this target: nothing is. The Stage-A telemetry arms (F = FREETOK reconstruction, T = UNION reconstruction) are struck pre-GPU: a host that computes p for rejection sampling can inject h_a itself with at most one least-squares solve against the cached lm_head (a 4096 x 128256 matvec plus a precomputed 4096 x 4096 inverse; this harness holds h_a directly through its final-norm catch). "Free" is D-5's convention; D-1's premise never priced shipping p or h_a to a disaggregated drafter (K-6 territory, not settled here). The oracle arm O (true h_a) is thereby the DEPLOYABLE free-actuator arm — SD^2-lite from the verifier's own distribution — whose V-vs-O readout sits outside the K-3 kill text and this lane's claim; it needs its own registration if run (label FREE-ACTUATOR WIN / NULL; a WIN is an SD^2 configuration result, not a sideband result). WHAT IT DOES NOT TOUCH, AND WHY: the existence results F-15/F-17 concern future-position concepts at offsets 2-3, which the anchor's p does not determine; but their free arm (a) is the same 657-summary proxy, so "beyond p?" arises there too, and the record's answer is the OQ-6 h-probe row of the same grids (margins_h_vs_a in results/real/probes_s1{,_replication}.json and probes_s3{,_replication}.json): under the identity a linear probe on h_a is a linear probe on the full logit vector, the strongest linear free-p baseline, and at the pinned L1 offset-2 cells it adds nothing over (a) — h-a margin S1 +0.0009 [-0.0087, +0.0107] primary, -0.0012 [-0.0112, +0.0090] replication; S3 -0.0026 [-0.0071, +0.0021], -0.0042 [-0.0093, +0.0015] — against c-a margins +0.0163 / +0.0141 (S1) and +0.0147 / +0.0173 (S3), and sits below (a) at offsets 3-4 on all four grids (offset-3 h-a -0.0205 / -0.0399 / -0.0111 / -0.0173). The K-2 margins are therefore not linear-logit information and F-15/F-17 stand; a NONLINEAR-in-p free baseline at the registered mlp256 capacity was never run and is the unrun K-2 comparator (not authorized here). The topic controls F-16/F-18 and the conversion nulls F-19/F-21 stand on their own. The only reconstruction target where telemetry could carry non-free information is h at FUTURE anchors (OQ-16's own anticipatory variant, never run); the fold-back-#2 hard stop forbids a third arc without the user's explicit exception. CLASSIFICATION: no partition branch is claimed. The limbo clause's condition is met (a Stage A never run after a PASS) and its rule, "neither limbo may ever be re-labeled a null", is honored; the catch-all is not invoked because the branch it would name (RECONSTRUCTION-CEILING NULL, "Stage-R leg 1 fails") is contradicted by the measured leg-1 pass; the disposition is the partition's own VOID precedent (no branch claimed, a dated note, the lane stops in practice under the fold-back-#2 hard stop). The lane is CLOSED BY DATED ERRATUM, mechanism TARGET-FREE; STAGE-R PASS stands as the mechanical outcome of its gate; the DR-3 OUTCOME's "reopens the Stage-A telemetry question" sentence is superseded. The lane ENDS. DISCLOSURE: the target was chosen for logging convenience (model.norm, one hook) rather than SD^2's mid-layer read (layers 3, L/2, L-2), which no identity of this kind ties to p (many-to-one; how much of those states p predicts is unmeasured); the free arm was built from the existing probe_input summaries rather than from D-5's definition; three diagnostics priced "beyond FREETOK's encoding" while the closed-form free comparator was one linear solve away. The observation came from the 2026-09-05 creative-lens adversarial pass over the gathered data (contrarian lens), reproduced by that pass's verifier, by an operator scratch check and by the committed tool; CRITIC RECOMPUTE ROUND (2026-09-05, before this commit): an independent implementation (float64 QR inverse; SVD least squares for the top-k solves; log-probabilities only) reproduced the forward numbers exactly, the full-logit inverse to 8 significant figures and every k < 4097 number to 7 or more digits; it exposed the first draft's k = 4097 row (median +0.9999, min 0.9988) as an artifact of an ill-conditioned normal-equations solve with a 1e-8 ridge, replaced by the committed SVD solve, and confirmed full column rank and the augmented-system conditioning above; it also blocked three sentences of the first draft — a "mechanical" catch-all classification against the limbo clause, an unargued "does not touch F-15/F-17", and a "standardized R^2" label — all rewritten before the commit. ATTESTATION: results/real/oq16_ha_identity.json carries git_commit 8542ef5-dirty; at run time the tracked dirt was this commit's own documentation and registration text, the tool and its test were untracked, and every input is sha-bound (lm_head shard sha256 and the twelve npz sha256s in the file).
    sideband/FINDINGS.md:267 · d147e8c0cd6b
  19. 2026-09-04sidebandfindings table row F-21dated note

    | F-21 | 2026-09-04 | OQ-15 Stage C CONVERSION-MODE NULL: telemetry head ties vanilla (+0.002664) and loses CI-clear to the free head (-0.081846 [-0.132484, -0.035567]) | stands (terminal for OQ-15); dated note 2026-10-10 (OQ-20, NON-CRITERION): …

    The full line, as written
    | F-21 | 2026-09-04 | OQ-15 Stage C CONVERSION-MODE NULL: telemetry head ties vanilla (+0.002664) and loses CI-clear to the free head (-0.081846 [-0.132484, -0.035567]) | stands (terminal for OQ-15); dated note 2026-10-10 (OQ-20, NON-CRITERION): replayed at seed 6 under the frozen policies the vs-vanilla leg reads -0.018029 [-0.065168, +0.028495] GUARD-HELD and the free leg (arm 7 minus arm 6) -0.053623 [-0.107526, -0.002448] GUARD-FREE-HELD; R2 MUTUAL-COVERAGE / REPRODUCES, R6 MUTUAL-COVERAGE / REPRODUCES (the differences' CI95 half-widths 0.0666 / 0.0707 above the +0.062 bar: a bar-sized move unresolved); one further seed; the verdict stands as computed |
    sideband/FINDINGS.md:58 · d147e8c0cd6b
  20. 2026-09-04sidebandfindings table row F-22other

    | F-22 | 2026-09-04 | OQ-16 Stage R PASS (linear reconstruction of the anchor hidden state h_a), magnitudes pathology-inflated; FREETOK (a free hashed encoding of p's top-32 token ids) beats TEL (the telemetry arm), UNION (TEL plus FREETOK) beats FREETOK +0.065180 [+0.062411, +0.067705]; "beyond p" unpriced after DR-4; …

    The full line, as written
    | F-22 | 2026-09-04 | OQ-16 Stage R PASS (linear reconstruction of the anchor hidden state h_a), magnitudes pathology-inflated; FREETOK (a free hashed encoding of p's top-32 token ids) beats TEL (the telemetry arm), UNION (TEL plus FREETOK) beats FREETOK +0.065180 [+0.062411, +0.067705]; "beyond p" unpriced after DR-4; Stage A frozen and unrun | qualified; superseded-in-part (dated notes inside the entry) |
    sideband/FINDINGS.md:59 · d147e8c0cd6b
  21. 2026-09-04sidebandF-22amendment

    DR-1 / DR-2 POST-GATE DIAGNOSTICS (registered with pinned labels and honest predictions AFTER the gate was read and BEFORE running; NON-CRITERION — they cannot reclassify the branch; instrument extension: …

    The full line, as written
    - DR-1 / DR-2 POST-GATE DIAGNOSTICS (registered with pinned labels and honest predictions AFTER the gate was read and BEFORE running; NON-CRITERION — they cannot reclassify the branch; instrument extension: a FEATURE-SUPPORT CLIP to the TRAIN [min, max] per feature, trivial-baseline rows and a REPORTED tel_vs_trivial in every run, and the equal-dims FREETOK arm = FREE ++ 2048-dim multi-hot of the anchor's top-32 next-token ids hashed mod 2048, deployment-free information from the same verify forward; the INVARIANCE LEG requires the extended tool at modes OFF to reproduce the registered numbers bit-exact). INVARIANCE LEG: holds on every gate-bearing field — all twelve deltas/CIs, the gate, per-population values, test MSEs, chosen lambdas, VAL tables, cosines, per-coordinate quartiles, and all 1,512 registered state rows are bit-identical; the REPORTED trivial MSE differs by ONE ULP (0.9900697257801255 registered vs 0.9900697257801256 extended; 1.1e-16 — the amendment accumulates trivial SSE per cluster before summing, the registered code kept a running scalar; the critic's own from-raws recompute had landed on ...259 for the same reason) and two of the three R^2 values (free, tel) inherit that last digit exactly (noise is bit-identical; both recompute bit-exact from the extended trivial). Explained per the pin; the diagnostics stand. The extended tool's tel_vs_trivial for the REGISTERED configuration reproduces the critic's number bit-exact: +0.10701294951277318 [-0.06983322605482828, +0.22087590283869973]; k3 +0.2163 [+0.2037, +0.2287]; oq15 -0.0002 [-0.3557, +0.2238]. DR-1 (support clip ON; pinned label): DEPLOYABLE-MARGIN CONFIRMED — tel_vs_trivial +0.250784 [+0.239316, +0.261616]; k3 +0.2484 [+0.2309, +0.2657]; oq15 +0.2532 [+0.2364, +0.2677] (the populations now agree — the pathology was the entire k3/oq15 asymmetry). Under the clip: TEST MSE free 0.9084 / tel 0.7393 / noise 0.9010 vs trivial 0.9901 -> R^2 +0.083 / +0.253 / +0.090; delta_noise +0.161692 [+0.153035, +0.170259] (stale points -0.000805 / -0.001043 / -0.000840 / -0.001285 / -0.000738 — the control tightens to sd 0.00022); delta_free +0.169079 [+0.160250, +0.177775]; gate label under the clip STAGE-R PASS (REPORTED — cannot reclassify); chosen lambdas free/tel 1e4 (INTERIOR, the first non-edge selections of the lane), noise 1e5 (edge); TEL centered-cosine median 0.502, per-coordinate R^2 quartiles [+0.208, +0.236, +0.271]. Grid edge under the clip is PARTIAL: free/tel/freetok live at 1e4 (interior) but noise-live and all 20 stale cells at 1e5 (lambda_grid_edge_selected true in the DR payload); the 3-sd bar under the clip is negative (-0.000274, all five stale points negative) — trivially cleared, noted. The registered run's magnitudes were pathology-DEFLATED for TEL (0.883 -> 0.739) and its baselines pathology-INFLATED; the honest picture is a linear R^2 of about a quarter of the anchor state from the address stream. Honest prediction (CONFIRMED expected): landed. DR-2 (FREETOK arm, equal dims, deployment-free; pinned label): TOKEN-IDENTITY-NOT-EXCLUDED — tel_vs_freetok = MSE(FREETOK) - MSE(TEL) = -0.004792 [-0.007192, -0.002600], CI-CLEAR NEGATIVE: the free next-token-identity arm reconstructs h_a slightly BETTER than telemetry (FREETOK test MSE 0.7345, R^2 +0.258, cosine median 0.513, per-coordinate quartiles [+0.219, +0.243, +0.274]; k3 -0.0066 [-0.0097, -0.0036]; oq15 -0.0030 [-0.0063, +0.0003]). freetok_vs_free +0.173871 [+0.164704, +0.182838] — a coarse hashed encoding of which 32 tokens p favors reaches a margin over FREE at least as large as telemetry's (delta_free under the clip +0.169). DESIGN LIMIT (critic blocker, wording corrected pre-commit): DR-2 compares TEL against FREETOK, not TEL++TOK against FREETOK, so it licenses "the telemetry margin is not evidence of information beyond what p already reveals at this readout" — NOT "the margin is composed of token identity" (FREETOK beats TEL in only 59/84 clusters; the two encodings could carry partly different information netting to a tie). DR-3 below is the registered resolver. freetok_vs_trivial +0.255576 [+0.244307, +0.266421]. Under stale targets TEL still edges FREETOK by ~+0.0098 (prompt-level slot statistics exceed prompt-level token statistics), so the block-aligned advantage of FREETOK over TEL is ~0.015. Honest prediction ("a large residual token share would be surprising"): the surprising side landed — a free encoding matches or exceeds the telemetry margin, recorded against the prediction. DR-3 ADDENDUM (dated 2026-09-04; registered pre-run in commit 2ec789f, run after; NON-CRITERION; pinned label): TELEMETRY-BEYOND-P — UNION (TEL ++ TOK, 4753 dims) beats FREETOK by union_vs_freetok +0.065180 [+0.062411, +0.067705] (k3 +0.0639 [+0.0597, +0.0680]; oq15 +0.0665 [+0.0629, +0.0697]) and beats its equal-dims alignment-destroyed anchor UNIONNOISE (NOISE ++ TOK) by +0.078089 [+0.075624, +0.080368] — both legs CI-clear on both populations. MAGNITUDE, carried as pinned: 37.5% of the free margin (freetok_vs_free +0.173871), ABOVE the pre-run "would surprise" line of a quarter — the surprising side landed a second time, recorded against the prediction. Under stale targets the union edges FREETOK by ~+0.0100 (prompt-level slot statistics); the stale-netted remainder ~+0.055 (heuristic — live and stale lambdas differ) leaves the +2048-dim dilution UNPAID and is a LOWER BOUND on the block-aligned addition, not the addition (critic blocker, corrected pre-commit); the dilution-priced estimate is the anchor readout union_vs_unionnoise +0.078 (conservative: its stale variants sit at -0.0026..-0.0030). Mutually consistent decomposition: block-aligned ~+0.081, prompt-level ~+0.010, dilution ~-0.026, net +0.065. Mirror: union_vs_tel +0.069971 [+0.067058, +0.072830] — p adds on top of telemetry about what telemetry adds on top of p; the two are overlapping but NOT redundant views. UNION R^2 +0.324 (test MSE 0.66931; cosine median 0.579; per-coordinate quartiles [+0.279, +0.307, +0.342]) vs FREETOK +0.258 and UNIONNOISE +0.245 (MSE(UNIONNOISE) - MSE(FREETOK) = +0.013 is dilution net of the shifted block's ~+0.013 prompt-level gain; gross dilution ~0.026 per +2048 dims here, ~5x F-16's scale — a calibration note for any Stage-A freeze). Lambdas interior at 1e4 for all arms but noise. The four shared arms reproduce the DR-1/DR-2 run bit-identically inside the six-arm invocation. WHAT THIS CORRECTS IN THE SECTIONS ABOVE AND BELOW: DR-2's licensed sentence ("not evidence of information beyond p") was the TEL-vs-FREETOK reading; DR-3 is the direct test — the S1 stream DOES carry block-aligned, dilution-priced information about the anchor state beyond FREETOK's 2048-bin hashed encoding of p's top-32 support, small in absolute terms (+0.065 standardized MSE) but CI-tight and replicating across two disjoint prompt populations. QUALIFIER (critic blocker): the anchor destroys all block-aligned slot information, so it cannot separate slot information that is a function of p's EXACT top-32 (identity lost to the ~62-ids-per-bin hash, value-to-id pairing lost) from information beyond p — "beyond p" is unpriced until an equal-dims better-token-encoding comparator (DR-4 candidate: FREE ++ TOK(hash A) ++ TOK(salted hash B) at 4753 dims, or a rank-paired TOK) is registered and run. The K-2 sentence stands with that qualifier: most of what telemetry reconstructs is a lossy copy of what p reveals; between a third and a half of its margin is not, depending on whether dilution is priced (union_vs_freetok / delta_free = 0.0652/0.1691 = 38.5% raw; union_vs_unionnoise / delta_free = 46% dilution-priced). WHAT IT DOES NOT LICENSE: any conversion-level claim — whether that R^2 increment (+0.066 of R^2; +0.065 of standardized MSE) buys acceptance at matched budget through an SD^2-style injection is Stage A's question, unmeasured, with two conversion nulls on this lane against it. DR-4 ADDENDUM (dated 2026-09-05; registered pre-run in commit c2181b1, run 10.9 s after; NON-CRITERION; the lane's LAST reconstruction-level diagnostic; pinned label; critic-recomputed bit-exact incl. a clean-worktree byte-identical re-run): COMPARATOR-TOOTHLESS — the equal-dims comparator FREETOK2 = FREE ++ TOK_A ++ TOK_B (2048 salted-hash bins valued by top_p; 4753 dims = UNION's) is CI-CLEAR WORSE than FREETOK at the registered readout: freetok2_vs_freetok -0.003085 [-0.004984, -0.001280] (k3 -0.003521 [-0.006788, -0.000549]; oq15 -0.002658 [-0.004983, -0.000447]; FREETOK2 wins 37/84 clusters). The teeth clause fails, so union_vs_freetok2 +0.068265 [+0.065783, +0.070556] (k3 +0.067390; oq15 +0.069122) is REPORTED, not labelling — identically union_vs_freetok minus freetok2_vs_freetok; it licenses one direction: at equal dims a value-paired second-hash encoding of p closes NONE of the DR-3 gap (union beats FREETOK2 in 84/84 clusters, min +0.036). FREETOK2 R^2 +0.255 vs FREETOK +0.258; freetok2_vs_free +0.170786 [+0.160740, +0.180599]; tel_vs_freetok2 -0.001707 [-0.004334, +0.000628] (straddle). Pinned magnitudes: 1.047 of the DR-3 margin (the surprising side (ii) fired mechanically and VACUOUSLY — the comparator is worse than FREETOK, so 'surviving' measures its loss) and -0.018 of the free margin. Honest prediction (teeth at +0.02..+0.05): FAILED ON SIGN, recorded against. CAVEAT (critic blocker; heuristic, not a registered readout): the teeth readout is UNANCHORED — no alignment-destroyed TOK_B twin was registered, so it charges the block its +2048-dim dilution; in the committed val_mse table FREETOK2 beats FREETOK at lambda 1e5 (0.77768 vs 0.79394) and the live-minus-stale-mean VAL gap is positive at every lambda (+0.047 at 1e-2 ... +0.031 at 1e4, +0.020 at 1e5), so the live -0.0031 is consistent with a block-aligned contribution of order +0.02..+0.03 offset by dilution of the same order at the selected lambda 1e4 — and equally with less. LICENSED: in a shared-lambda ridge TOK_B does not net a positive margin over FREETOK at either arm's VAL-optimal lambda; its block-aligned contribution is not separable from dilution under the registered readout. NOT licensed: 'extracts nothing', 'saturated', 'pure dilution' (the stale variants sit at lambda 1e5, where dilution is an order smaller). The anchored test (FREETOK ++ alignment-destroyed TOK_B) would resolve it but is a DR-5 under the fold-back-#2 hard stop — run only if the user authorizes an exception in those words. Dated erratum to the registration: TOK_B is sparse like TOK_A, with heavy-tailed values (median nonzero 0.022; 4.7% float16 underflow), not 'dense-valued'. The six DR-3 arms reproduce bit-identically (3108/3108 shared state rows); the -dirty meta is attested moot by the clean-worktree re-run. WHAT IT MEANS: DR-4 did NOT deliver the basis the registration said a Stage-A telemetry test NEEDS; 'beyond p' stays UNPRICED; the pre-registered recommendation stands (treat as NOT confirmed for the Stage-A decision) and Stage A, if run, proceeds on the user's decision with that on the table; FREETOK stays the Stage-A free comparator; a Stage-A telemetry WIN could only ever be licensed as 'converts beyond FREETOK, the strongest free encoding tested, at the actuator', never 'beyond p'.
    sideband/FINDINGS.md:260 · d147e8c0cd6b
  22. 2026-09-04sidebandF-22dated note

    WHAT OQ-16 ACTUALLY ANSWERED (the ledger's literal question — "what fraction of SD^2's vector survives reconstruction from the address stream?"): linearly, about a quarter (R^2 ~0.25, median direction cosine ~0.50 in standardized space, with the support clip) — …

    The full line, as written
    - WHAT OQ-16 ACTUALLY ANSWERED (the ledger's literal question — "what fraction of SD^2's vector survives reconstruction from the address stream?"): linearly, about a quarter (R^2 ~0.25, median direction cosine ~0.50 in standardized space, with the support clip) — and a free encoding of the output distribution's top-32 token identities, which the same verify forward hands over anyway, recovers at least as much (CI-clear). At this readout the telemetry margin is therefore not evidence of information beyond what p already reveals; whether telemetry adds anything ON TOP of p is DR-3's question. (DATED NOTE 2026-09-04: DR-3 answered it — see the DR-3 ADDENDUM in the DR block above; this sentence, the "ties a free encoding of p" phrase, and the K-2 location below are SUPERSEDED in the addendum's qualified sense: the stream adds beyond FREETOK's hashed encoding of p, CI-clear; "beyond p" itself is unpriced pending DR-4.) Located against the kill criteria: this is K-2's "a lossy copy of information the decoder already has" at the reconstruction level; it does not re-fire K-2 (that verdict was on future-concept probes at offsets >= 2, F-15, a different target). This is the OQ-6 h-probe story from the other side: extracting FROM h under a cheap budget lost to telemetry (committed), and reconstructing TOWARD h from telemetry ties a free encoding of p. - FORMAL STATE + STAGE-A DECISION INPUTS: by the pinned partition, Stage R is a PASS (the registered FREE arm was probe_input's 657 dims; FREETOK was not a registered baseline and DR-2 is NON-CRITERION) and Stage A (SD^2-style draft-side injection of the reconstructed vector) is AUTHORIZED and USER-GATED; nothing runs from this finding, and the lane hard-stops after Stage A regardless of its outcome. HONEST WEIGHT for that decision, in the registration's own words: DR-2's TOKEN-IDENTITY-NOT-EXCLUDED label is carried into the Stage-A cost decision and mandates FREETOK (or a stronger token-identity encoding) as a free comparator in any actuator economics. With DR-3 in view (the addendum above): a Stage-A telemetry test now HAS a reconstruction-level basis — the UNION reconstruction is CI-clear better than the FREETOK one (R^2 +0.324 vs +0.258), so the Stage-A telemetry claim pinned for the freeze is UNION-vs-FREETOK at the actuator, both free-augmented (the STAGE-A FREEZE landed 2026-09-05 in docs/BLUEPRINT.md, pre-GPU, critic-gated twice: SD^2-style up-projection injection with a frozen drafter and a rank-512 zero-initialised adapter; arms vanilla / oracle_h (TRUE h_a, the actuator ceiling, D-2-named) / free_inject = FREETOK / tel_inject = UNION on the fresh 210-prompt data/prompts_oq16_verdict.jsonl population (mechanical paraphrase refusal against all four committed sets, 0 overlaps); ladder A0 (oracle vs vanilla at the unchanged E6 +0.062 bar) before A1 (the free-actuator readout F-vs-V and the telemetry verdict T-vs-V AND T-vs-F); each rung user-launched; DR-4's COMPARATOR-TOOTHLESS label and its pinned recommendation are carried into the user's A1 decision); a TEL-vs-FREETOK actuator test would have none (TEL is CI-clear slightly worse than FREETOK). Two conversion nulls on this lane (F-20/F-21) remain the closest precedents, and whether a +0.065 standardized-MSE increment in the reconstructed vector moves acceptance at matched budget is unmeasured — no probability is licensed either way. What Stage A could also measure is whether an SD^2-style injection of a FREE reconstruction converts at all (a different question, outside the K-3 kill text's telemetry claim, and not authorized by this partition without its own registration). Mandatory Stage-A freeze elements pinned by this finding: the support clip (or an sd floor) on every live feature — an actuator injecting a 5.76-MSE reconstruction on an out-of-support prompt is worse than nothing (critic A3); a lambda family containing the trivial predictor; the pairwise-selected-budget check (F-21); and — DR-2 having landed TOKEN-IDENTITY-NOT-EXCLUDED — the FREETOK arm (or a stronger token-identity encoding) as a mandatory free comparator in the actuator economics. What the pass says: a linear map from the address stream recovers the anchor state's direction at median cosine ~0.5 in standardized space and beats the free features in every test cluster; what it does not say: that the recovered subspace is the steering-relevant one (Stage A's question), that the margin survives raw-h-space injection (the standardized/raw distinction is pinned), or that any of this converts to tokens/call at matched budget (two nulls on this lane say the conversion bar is high).
    sideband/FINDINGS.md:261 · d147e8c0cd6b
  23. 2026-09-03sidebandfindings table row F-19erratum

    | F-19 | 2026-09-03 | K-3 Attempt #1 NULL: telemetry block scheduling +0.006860 [-0.015024, +0.025572] tokens/call vs a +0.062 bar | stands (K-3 fired); erratum-noted; dated note 2026-10-04 (F-27): …

    The full line, as written
    | F-19 | 2026-09-03 | K-3 Attempt #1 NULL: telemetry block scheduling +0.006860 [-0.015024, +0.025572] tokens/call vs a +0.062 bar | stands (K-3 fired); erratum-noted; dated note 2026-10-04 (F-27): the val-selected +0.059725 sits at the 0.8840 percentile of the n = 30 cross-seed seed-noise null on this population (exceeded by 0.1160 of pooled draws), a single draw; the val comparison was same-seed (seed 0), different-policy, on a different prompt set and a max over 32 search evals, none modeled; dated note 2026-10-10 (OQ-20, NON-CRITERION): replayed at seed 6 under the frozen policies the vs-vanilla leg reads +0.010340 [-0.009288, +0.030521] GUARD-HELD and the free leg (arm 4 minus arm 3) +0.010109 [-0.009724, +0.030422] GUARD-FREE-HELD; R2 MUTUAL-COVERAGE / REPRODUCES, R6 MUTUAL-COVERAGE / REPRODUCES (the free leg nearly the vs-vanilla leg, abs(R5 - R1) = 0.000231: not two confirmations); one further seed; the verdict stands as computed |
    sideband/FINDINGS.md:56 · d147e8c0cd6b
  24. 2026-09-02sidebandF-19amendment

    KILL TEXT: this is K-3 firing exactly as registered — probe margins (F-15/F-17 existence, both real) did not convert to acceptance-side throughput through this actuator at this pair/scale.

    The full line, as written
    - KILL TEXT: this is K-3 firing exactly as registered — probe margins (F-15/F-17 existence, both real) did not convert to acceptance-side throughput through this actuator at this pair/scale. Per the FOLD-BACK COUNT AMENDMENT (2026-09-02, critic-gated; it WIDENED the kill text's original single fold-back to the structured two-lane sequence, a dated pre-verdict-read change disclosed in its own text): fold-back #1 = OQ-15 (outcome-supervised acceptance head; its arc registration must pin the exhaustive terminal-outcome partition before running); an information-ceiling null there skips OQ-16 and stops the lane; a conversion-mode null hands off to OQ-16 (vector reconstruction -> SD^2 injection); hard stop after #2 or the skip. SALVAGE per the kill doc: the per-offset gating rules and staleness map stand; the byproduct logs (per-block telemetry + outcomes + anchor states + analytic acceptance, ALL arms, BOTH seeds, ~10 runs x 210 prompts) are the OQ-15/OQ-16 training data, collected for free by the registered amendment. No wall-clock claims (D-9); no acceptance-rate claims beyond the labeled tokens/call estimand.
    sideband/FINDINGS.md:222 · d147e8c0cd6b
  25. 2026-09-01sidebandfindings table row F-17erratum

    | F-17 | 2026-09-01 | K-2 survived at S3: MoE router expert ids beat the free baseline at offset 2, pinned replication +0.0173 [+0.0133, +0.0221] | stands against the 657-summary PROXY; DATED ERRATUM 2026-09-06 (F-24): …

    The full line, as written
    | F-17 | 2026-09-01 | K-2 survived at S3: MoE router expert ids beat the free baseline at offset 2, pinned replication +0.0173 [+0.0133, +0.0221] | stands against the 657-summary PROXY; DATED ERRATUM 2026-09-06 (F-24): ABSORBED by the full free baseline at the registered reader at the pinned cells L1 offset 2, TWIN-ROBUST at the corrected twin; mechanism qualified by F-18; DATED 2026-09-14 (F-25, OQ-18 at the registered reader): L2 ABSORBED (F-25); L3 SPLIT at the L3 group lens (replication fired; F-25); DATED LENS NOTE 2026-10-05 (F-28, STAGE LENS; OQ-13 (ii)): through the BALANCED L3 lens (16 groups of 31 or 32, results/real/concept_groups_olmoe_balanced_v1.json) the (c) - (a) linear off2 margin reads +0.0221 [+0.0187, +0.0258] / +0.0195 [+0.0160, +0.0232], 0.307 / 0.267 of the recorded +0.0721 / +0.0730 — LENS-SURVIVES (the margin exists under a size-capped partition); against the committed L1 (+0.0147 / +0.0173) and L2 (+0.0206 / +0.0207) margins the growth over L2 is NOT-SETTLED on both harvests (combined L3-GROWTH-NOT-SETTLED; P's CI contains L2's margin; not like-for-like across label spaces), so the "coarser claim" sentence is not written; off3 (+0.0419 / +0.0364) and off4 (+0.0207 / +0.0227) are not re-read at the balanced lens; the recorded-lens number stands regardless |
    sideband/FINDINGS.md:54 · d147e8c0cd6b
  26. 2026-09-01sidebandfindings table row F-18erratum

    | F-18 | 2026-09-01 | The S3 margin is not prefix-topic state at the histogram level; about 1/10 of the L3 margin is topic-shaped | stands; erratum-noted; dated note 2026-09-06 (F-24): the mechanism reading is moot for the S3 L1 off2 margin at the registered reader; the L3 statements are untouched; dated note 2026-09-14 (F-25): …

    The full line, as written
    | F-18 | 2026-09-01 | The S3 margin is not prefix-topic state at the histogram level; about 1/10 of the L3 margin is topic-shaped | stands; erratum-noted; dated note 2026-09-06 (F-24): the mechanism reading is moot for the S3 L1 off2 margin at the registered reader; the L3 statements are untouched; dated note 2026-09-14 (F-25): mechanism reading moot at L2 at the registered reader |
    sideband/FINDINGS.md:55 · d147e8c0cd6b
  27. 2026-09-01sidebandF-18erratum

    caveats: NON-CRITERION: this prices F-17's signal; it cannot amend F-17's verdict. "M2/M2n fire 6/6" is true ONLY under a CI-positive reading — 5/6 rows clear the +0.005 floor in each family, and L1/off3 is CI-positive-immaterial in both (M2 +0.002069 [+0.000503, +0.003900]; M2n +0.002631 [+0.001169, +0.004487]) — …

    The full line, as written
    - caveats: NON-CRITERION: this prices F-17's signal; it cannot amend F-17's verdict. "M2/M2n fire 6/6" is true ONLY under a CI-positive reading — 5/6 rows clear the +0.005 floor in each family, and L1/off3 is CI-positive-immaterial in both (M2 +0.002069 [+0.000503, +0.003900]; M2n +0.002631 [+0.001169, +0.004487]) — F-17's L1/off3 profile again; the replication M2n at L1/off3 is the one non-reproducing row, CI-crossing with one negative seed (+0.000883 [-0.002524, +0.003312], per-seed +0.00201/-0.00030/+0.00094) — the weaker tail, disclosed as at F-17. DATED ERRATUM (2026-09-01) to the instantiation block's disclosure (3), which reads "Dims differ from S1's 560/3266": the recorded t dim is 560, IDENTICAL to S1 (the OLMoE alphabets merge to the same 511/16); only the slots block differs (1024 vs S1's 2048), and the "3266" half inherits F-16's own 3266->3265 erratum — the block text stands, corrected here per the erratum convention. The payloads' registered/conclusion.rule strings date the P1 template (2026-08-31), not this instantiation (2026-09-01) — the same cosmetic residue class F-17 disclosed for decision.rule. The (atn) reading carries VERBATIM from F-16: the shifted-slots arm could in principle carry aligned-equivalent signal, a channel whose bias could only make M2n an UNDERestimate — SURVIVES rides on M2 and is unaffected. PROMPT-TOKEN RESIDUAL as registered: (t) omits prompt tokens (the harvest stores generated tokens only); a failure would have deflated a fortiori, survival leaves the prompt-augmented prefix control as the formally open residual and the cheapest next experiment if mechanism language is ever contested. HISTOGRAM FAMILY ONLY: richer prefix representations (sequence models over the prefix, k-NN/embedding similarity) fall under the residual, not this rejection; any richer control is a new registration. Offsets {2,3} ran verbatim per the template although F-17 establishes L1 signal at off2 only — off3 rows reported, no cell swap, conclusion weight on the pinned primary cell. Linear probes only (registered economy); 30-cluster power; L3 mega-group lens (two groups carry ~461/510 of the concept mass — F-17 carry-forward); glue-majority floors compress all margins in every space (F-17 carry-forward). No acceptance or wall-clock claims; K-3 untouched — this says the S3 signal is real and not derivable from emitted tokens at this family level, not that it converts.
    sideband/FINDINGS.md:208 · d147e8c0cd6b

August 20262 entries

  1. 2026-08-31sidebandfindings table row F-15erratum

    | F-15 | 2026-08-31 | K-2 survived at S1: weight-slot telemetry beats the free baseline at offsets 2-3, pinned replication +0.0141 [+0.0081, +0.0203] | stands against the 657-summary PROXY; DATED ERRATUM 2026-09-06 (F-24): …

    The full line, as written
    | F-15 | 2026-08-31 | K-2 survived at S1: weight-slot telemetry beats the free baseline at offsets 2-3, pinned replication +0.0141 [+0.0081, +0.0203] | stands against the 657-summary PROXY; DATED ERRATUM 2026-09-06 (F-24): ABSORBED by the full free baseline at the registered reader at the pinned cells L1 offsets 2-3, TWIN-ROBUST at the corrected twin; mechanism qualified by F-16; DATED 2026-09-14 (F-25, OQ-18 at the registered reader): L2 ABSORBED (F-25); L3 NOT-SHOWN at the L3 group lens (F-25) |
    sideband/FINDINGS.md:52 · d147e8c0cd6b
  2. 2026-08-31sidebandfindings table row F-16erratum

    | F-16 | 2026-08-31 | The S1 margin is not prefix-topic state at the histogram level; passes the equal-dims noise control | stands; erratum-noted; dated note 2026-09-06 (F-24): the shift-17 equal-dims twin is NOT inert under a null direction, and the mechanism reading is moot for the S1 L1 off2-3 margin at the registered reader; …

    The full line, as written
    | F-16 | 2026-08-31 | The S1 margin is not prefix-topic state at the histogram level; passes the equal-dims noise control | stands; erratum-noted; dated note 2026-09-06 (F-24): the shift-17 equal-dims twin is NOT inert under a null direction, and the mechanism reading is moot for the S1 L1 off2-3 margin at the registered reader; dated note 2026-09-14 (F-25): mechanism reading moot at L2 at the registered reader |
    sideband/FINDINGS.md:53 · d147e8c0cd6b

Lessons from the record's own limitations

sideband/docs/LESSONS.md:1 · 61a765b6939d
In plain wordsA note the sideband record wrote about its own limitations after a read only audit of its own code and record. For each one: the mechanism of the mistake, what it cost as measured, the rule the next instrument follows, and where the repository itself does not yet follow that rule.

Written 2026-10-16 as the closing step of lane B of the public-cut plan, after the 2026-10-15 image audit (docs/audits/2026-10-15-image-audit/README.md, the audit's report with the recompute scripts beside it). sideband/docs/LESSONS.md:3 · 61a765b6939d

The note's preamble, as written

Written 2026-10-16 as the closing step of lane B of the public-cut plan, after the 2026-10-15 image audit (docs/audits/2026-10-15-image-audit/README.md, the audit's report with the recompute scripts beside it). Each entry below is a limitation that stays in the evidence: the instruments are frozen (their bytes are attested by committed payloads) and the registered runs are not repeated, so the record carries a dated note or erratum beside the affected text instead of a fix. For each one: the mechanism, the cost as measured, the rule the next instrument follows, and where the repository itself does not yet follow the rule. The dated notes in FINDINGS.md, DECISIONS.md and docs/BLUEPRINT.md are the primary text; this file is a derived reading aid, and on any disagreement the record's dated note wins. Figures quoted from the audit folder's reruns of 2026-10-16 name their file; the counts in entries 7, 8 and 13 were measured on the tree the same day by the commands named. sideband/docs/LESSONS.md:3 · 61a765b6939d

1. A mechanism sentence copied forward past the code

sideband/docs/LESSONS.md:15 · 61a765b6939d
Mechanism
"the HF adapter recomputes prefixes" was written on 2026-08-29 and was false from a commit later that day (b1dae57, cropped KV caches by default). Twelve files carried the sentence, eight of them written after that commit, because each new document copied it from the last one instead of the code; a thirteenth, docs/METHOD.md, quotes the critic's rule that carried it. sideband/docs/LESSONS.md:17 · 61a765b6939d
Measured cost
fifteen false or stale statements, zero numbers moved; the ban itself was right for a different reason (no benchmark harness exists). sideband/docs/LESSONS.md:22 · 61a765b6939d
Rule
a mechanism claim has one home (here DECISIONS.md D-9 and its dated note of 2026-10-16); every other file cites the home rather than restating the mechanism, and a ban states its own ground, not a code detail that can change. sideband/docs/LESSONS.md:24 · 61a765b6939d
Not yet followed by
this lane's own dated notes, which restate the b1dae57 mechanism beside each stale sentence so that a reader of that sentence alone sees the correction; the cost is accepted, and tests/test_record_claims.py pins the code default they cite. sideband/docs/LESSONS.md:27 · 61a765b6939d

Cites D-9

2. Intervals that are conditional on the trained seeds

sideband/docs/LESSONS.md:31 · 61a765b6939d
Mechanism
every probe CI is a prompt-cluster bootstrap over counts summed across three train seeds (tools/run_probes.py, frozen), so it is conditional on those three trained probes; the registered all-seeds rule protects the fire direction only. The registered disclosure that the seed "only shuffles AdamW batch order" was false: the trainer seeds the initial weights too (dated erratum in docs/BLUEPRINT.md). sideband/docs/LESSONS.md:33 · 61a765b6939d
Measured cost
across the 40 D rows the seed SE runs 0.33 to 3.26 times the bootstrap SE (docs/audits/2026-10-15-image-audit/seed_var_out.json, which brackets 72 rows of F-24, its corrected twin and F-25; F-15 to F-18 and F-28 are not bracketed). No label moves under the three approximations. Three bracketed readings are seed-sensitive: F-25's S1 L2 deciding row, +0.0022 [+0.0003, +0.0039], reads [-0.0012, +0.0055] / [-0.00065, +0.00498] / [-0.0002, +0.0064]; F-24's S1 corrected-twin CI-clear reading is lost under two of three; F-24's S3 linear CI-positive qualifier is lost under all three. A fourth, descriptive reading, the primary S3 L3 Mn_r below-zero value behind F-25's post-hoc "same shape" sentence, has its upper above zero under all three. Three seeds are two degrees of freedom; the component is bracketed, not measured. sideband/docs/LESSONS.md:38 · 61a765b6939d
Rule
a registration states what its interval conditions on, in the regime line, and either registers a seed-inclusive interval beside the cluster bootstrap or says the seed component is unmeasured; a disclosure about a trainer is written after reading the trainer. sideband/docs/LESSONS.md:48 · 61a765b6939d
Not yet followed by
F-28's SURVIVES-STRONGER cells and the F-15 to F-18 margins, which carry the conditioning note but no bracket (F-28's dated note of 2026-10-16 says so). sideband/docs/LESSONS.md:52 · 61a765b6939d

Cites F-24, F-25, F-15, F-18, F-28, F-15 to F-18

3. A scorer that ran live without its support clip

sideband/docs/LESSONS.md:55 · 61a765b6939d
Mechanism
the frozen Stage-C heads (tools/train_outcome_head.py) export TRAIN moments but no TRAIN support, and the live scorer applies only the [0, 1] output clip; the standardization pathology that F-22 (iii) records for Stage R reached the live decisions of F-21, where the record said nothing about it. Stage A was pinned with "the support clip live" without back-checking Stage C. sideband/docs/LESSONS.md:57 · 61a765b6939d
Measured cost
a counterfactual TRAIN-support clip changes 41 of 17,185 consulted decisions on arm 6 and 41 of 17,616 on arm 7 (0.24% / 0.23%); a first-order estimate of the direct tokens/call effect is at most 0.0119 / 0.0116 (five drafted tokens between a stop at position 1 and an extension to six; the rest of a changed prompt's trajectory is not priced), inside every interval of F-21; the replay matches all logged decisions (docs/audits/2026-10-15-image-audit/replay_c1.json). sideband/docs/LESSONS.md:62 · 61a765b6939d
Rule
a scorer that runs live carries its TRAIN support clip, pinned at the freeze; when a pathology is found in one stage, every stage that shares the code path is back-checked in the same landing. sideband/docs/LESSONS.md:69 · 61a765b6939d

Cites F-22, F-21

4. The percentile bootstrap at thirty clusters

sideband/docs/LESSONS.md:73 · 61a765b6939d
Mechanism
a percentile cluster bootstrap narrows by (n-1)/n and uses z where t belongs, so it under-covers at small cluster counts; src/sideband/stats.py called it "the conservative end" (relative to Clopper-Pearson), and no document disclosed the shortfall at the record's cluster counts. sideband/docs/LESSONS.md:75 · 61a765b6939d
Measured cost
coverage 0.897 / 0.939 / 0.948 at 10 / 30 / 60 clusters for a nominal 0.95 on an F-24-shaped paired simulation of paired_cluster_delta_ci (docs/audits/2026-10-15-image-audit/coverage_out.json). The nearest label-deciding bound to the +0.005 floor is F-28's RN at L1 replication gate upper +0.004951 (results/real/readers_s3_r.json), which a t-for-z correction carries to +0.005026: under that correction the cell would read NOT-SHOWN instead of ABSORBED by the frozen rule, while the registered estimator is the percentile bootstrap, the label stands as registered, and the final S3 word K-2-ISOLATED does not change (F-28's dated note). sideband/docs/LESSONS.md:79 · 61a765b6939d
Rule
METHOD states the coverage of the interval method at the cluster count in use before the first landing that uses it; a new instrument at thirty clusters uses a t-multiplier or a bias-corrected bootstrap and says which. sideband/docs/LESSONS.md:87 · 61a765b6939d
Not yet followed by
every landed interval, by construction; the record discloses the one cell the correction would move and leaves the labels as registered. sideband/docs/LESSONS.md:90 · 61a765b6939d

Cites F-24, F-28, K-2

5. A sign sentence from one seed

sideband/docs/LESSONS.md:93 · 61a765b6939d
Mechanism
the synthetic experiments ran one seed per curve (e04: seed 200), and F-4 wrote "net harmful at every offset" from that curve at rho 0.6. sideband/docs/LESSONS.md:95 · 61a765b6939d
Measured cost
over seeds 200 to 209 the offset-1 value is +0.0005 on average (SE 0.0010), negative on 5 of 10, with the committed seed the most negative of the ten; offsets 2 to 6 are negative on all ten, so the rule of thumb stands and one committed clause does not (docs/audits/2026-10-15-image-audit/f4_seeds_out.json). sideband/docs/LESSONS.md:97 · 61a765b6939d
Rule
a sign sentence is written only after the seeds that could overturn it have run; a curve from one seed is labelled as one seed in the finding's regime line (F-4's regime line carries that clause since 2026-10-16; F-3 and F-5 are one-seed curves too). sideband/docs/LESSONS.md:101 · 61a765b6939d

Cites F-4, F-3, F-5

6. A ledger that omits a term that differs between arms

sideband/docs/LESSONS.md:105 · 61a765b6939d
Mechanism
F-9's bit ledger priced corrections and hints but not the bonus token the verifier emits when a whole block is accepted, and the omission was not listed among the ledger's caveats; the bonus rate differs between arms, so the omission favours the arm it was meant to judge. sideband/docs/LESSONS.md:107 · 61a765b6939d
Measured cost
with the bonus term priced the dense hint's break-even budget of 0.162 bits per token is about 26% high and the loss ratio reads about 26:1 instead of 20:1; the direction holds (docs/audits/2026-10-15-image-audit/e07_ledger_out.json). sideband/docs/LESSONS.md:111 · 61a765b6939d
Rule
a ledger lists every term it omits and says whether each is arm-invariant; a term that is not arm-invariant is priced or no budget number is stated. sideband/docs/LESSONS.md:115 · 61a765b6939d

Cites F-9

7. Counts and dates in prose

sideband/docs/LESSONS.md:118 · 61a765b6939d
Mechanism
counts were typed on a date and never re-measured: 30 path-bearing files (80 at HEAD), 609 and 711 tests (734), 27 findings in the lint line (29), 45 test files (47), 40 top-level tools (41), "batch-16" in a payload for a batch-10 run, "smoke runs no longer overwrite evidence" when each script's own flag still does, and the tool-docstring counts of docs/CONVENTIONS.md. sideband/docs/LESSONS.md:120 · 61a765b6939d
Measured cost
seven stale statements a reader can falsify with one command each; no number of the record moved. This lane's first draft typed three more ("fourteen files", a helper count that did not add up, "two path kinds"), caught by its reviewers. sideband/docs/LESSONS.md:125 · 61a765b6939d
Rule
every count in prose carries its date and the command that measured it, is written into the file by the generator that measured it (never retyped from a print), and is re-measured whenever the quantity it counts can have changed, not only when the file is touched. sideband/docs/LESSONS.md:128 · 61a765b6939d
Not yet followed by
the counts in this file and in README.md, which carry their date but not their commands; and the next landing that adds a test or a tool moves 734 and 41. sideband/docs/LESSONS.md:132 · 61a765b6939d

8. One provenance stamp with two meanings

sideband/docs/LESSONS.md:135 · 61a765b6939d
Mechanism
the instruments' git helpers disagree on what -dirty means. Measured at HEAD on 2026-10-16: in tools/run_*.py twelve copies count modified tracked files only and four count untracked files too (run_probes, run_posthoc_topic, run_k3_arms, run_error_predictor); the adapter hf_specdec, the harvesters and the trainers count untracked files, 20 plain git status --porcelain call sites against 13 with --untracked-files=no under src/, tools/ and experiments/ (the harness's tools/cc excluded); two documents each described one of the behaviours as the rule. sideband/docs/LESSONS.md:137 · 61a765b6939d
Measured cost
a provenance reader gets contradictory definitions; nothing is unattested, but what a stamp attests must be read from the instrument's own helper. sideband/docs/LESSONS.md:144 · 61a765b6939d
Rule
one provenance helper in the library with one definition; until a frozen instrument's copy is retired, the record names which definition each payload carries. sideband/docs/LESSONS.md:146 · 61a765b6939d
Not yet followed by
every instrument and the adapter; the dated notes of 2026-10-16 in docs/SHARP_EDGES.md and docs/REPORT.md give the split. sideband/docs/LESSONS.md:148 · 61a765b6939d

9. A bold claim without a forward pointer

sideband/docs/LESSONS.md:151 · 61a765b6939d
Mechanism
the registration log announced EXISTENCE CONFIRMED in bold at two sites and in two verdict lines, and later read ABSORBED and K-2 FIRED thousands of lines below; nothing at the claims pointed forward, so a top-down reader or a quotation of the bold line gives the opposite of the record's final state. sideband/docs/LESSONS.md:153 · 61a765b6939d
Measured cost
no number, but the one defect both walkthrough readers of the audit hit first; the premise paragraph and docs/REPORT.md's abstract carried the same exposure. sideband/docs/LESSONS.md:157 · 61a765b6939d
Rule
when a later entry overturns or narrows a bold claim, the landing that records the later entry appends a dated forward pointer at the claim in the same commit; the registered text is never rewritten. sideband/docs/LESSONS.md:159 · 61a765b6939d
Not yet followed by
CITATION.cff's abstract, which still reads "existence is shown" (the export tool rewrites that file's authors block, so its text is lane C's to touch). sideband/docs/LESSONS.md:162 · 61a765b6939d

Cites K-2

10. Undefined voices

sideband/docs/LESSONS.md:165 · 61a765b6939d
Mechanism
the registration log scored predictions as "the assistant's" against "the critic's" and reported "critic-recomputed to 0.0" without saying that the assistant is an AI co-author, the critic an LLM review agent, and the user, operator and owner one person. sideband/docs/LESSONS.md:167 · 61a765b6939d
Measured cost
the review chain read as more independent than it is, to every reader who did not already know the harness. sideband/docs/LESSONS.md:171 · 61a765b6939d
Rule
a document that names roles defines them in its front matter; an LLM review is named as such wherever its recompute is cited. sideband/docs/LESSONS.md:173 · 61a765b6939d

11. A threshold supported by the run it then judged

sideband/docs/LESSONS.md:176 · 61a765b6939d
Mechanism
the R1 fidelity rider's entropy tolerance cited "v2 max 0.0773" among its support, and v2 was the rerun then judged against that tolerance; both payloads arrived in one commit and the text never defined v2. sideband/docs/LESSONS.md:178 · 61a765b6939d
Measured cost
the entropy leg's pass is not independent of its threshold; the overlap and mass legs are independent and stand, and the AND-structure limits what the entropy leg alone licenses (dated erratum in docs/BLUEPRINT.md). sideband/docs/LESSONS.md:181 · 61a765b6939d
Rule
calibration data and the judged run are different draws, and the registration says which draw supports each tolerance. sideband/docs/LESSONS.md:184 · 61a765b6939d

12. Citations recalled from memory

sideband/docs/LESSONS.md:187 · 61a765b6939d
Mechanism
an attribute of a benchmark ("seeded realizations") entered registered text from a recalled idea list; the landing's source check found no such thing, and the correction sat far from the claim. sideband/docs/LESSONS.md:189 · 61a765b6939d
Measured cost
one false attribute in registered text, corrected in its OUTCOME; other recalled identifiers remain marked "to verify" in docs/PRIOR_ART.md. sideband/docs/LESSONS.md:192 · 61a765b6939d
Rule
an identifier or attribute recalled rather than read is marked "recalled, unverified" at the point of use until it is read at the source; a correction is placed beside the claim, not only in the outcome. sideband/docs/LESSONS.md:194 · 61a765b6939d
Not yet followed by
the recalled identifiers docs/PRIOR_ART.md lists for verification, a lane of their own. sideband/docs/LESSONS.md:197 · 61a765b6939d

13. Absolute paths stored in evidence

sideband/docs/LESSONS.md:200 · 61a765b6939d
Mechanism
instruments recorded the operator's absolute invocation paths in payloads, and three recorded path kinds are read back: the Phase-3 companion paths as basenames by the frozen reader-lens tool with the running OS's semantics, the amendment's rider path by one test, and a relative backslash harvest_dir by one prefetch test. sideband/docs/LESSONS.md:202 · 61a765b6939d
Measured cost
six tests red on a fresh Linux clone (three, two and one) until lane A (c41f9f1, 2026-10-16) handled them test-side; 80 tracked payloads under results/ and 87 tracked files tree-wide carry such paths (git grep -l -E on the home-path pattern) and are disclosed, not edited. sideband/docs/LESSONS.md:206 · 61a765b6939d
Rule
an instrument stores repository-relative paths; a reader that must consume a recorded path normalises it with PureWindowsPath before taking a name. sideband/docs/LESSONS.md:210 · 61a765b6939d
Not yet followed by
the instruments, whose shared guard does not relativize paths, so a new run still records the operator's paths (README.md's provenance note says so). sideband/docs/LESSONS.md:212 · 61a765b6939d

14. A smoke flag that writes into the evidence

sideband/docs/LESSONS.md:215 · 61a765b6939d
Mechanism
the driver's --smoke redirects to a gitignored folder, but each experiment script's own --smoke does not, and the release notes generalised the driver's fix. sideband/docs/LESSONS.md:217 · 61a765b6939d
Measured cost
a one-command rerun of a script dirties a committed payload (caught by the evidence-bytes test and git status; no committed number changed). sideband/docs/LESSONS.md:219 · 61a765b6939d
Rule
a smoke path redirects at the script level, and the document that claims a fix names the exact entry point it covers. sideband/docs/LESSONS.md:221 · 61a765b6939d
Not yet followed by
experiments/e01 to e07, lane C's fix. sideband/docs/LESSONS.md:223 · 61a765b6939d

15. The audit method that found these

sideband/docs/LESSONS.md:225 · 61a765b6939d
Mechanism
twenty-four finders read the code in two modes, a tone reader read the registration log, two walkthrough readers used the export, and twelve refuters re-derived every number with the default set to refute. sideband/docs/LESSONS.md:227 · 61a765b6939d
Measured cost
of 375 findings, 45 were refuted and 136 downgraded; the raw finder list would have misled by about half, and zero committed labels were overturned. sideband/docs/LESSONS.md:230 · 61a765b6939d
Rule
a finding is acted on after a refuter has recomputed it; a review that cannot be recomputed is labelled judgement, as the tone section of the audit report is. sideband/docs/LESSONS.md:232 · 61a765b6939d
Not yet followed by
this lane itself, which acted on B1 (verified by the orchestrator's own read and re-verified here, no refuter) and on B7 (tone, judgement, no refuter), and says so; the lane's own three reviewers then found eleven wording and count errors in its first draft, which is the rule's point. sideband/docs/LESSONS.md:234 · 61a765b6939d

16. The site survey before registration

sideband/docs/LESSONS.md:239 · 61a765b6939d
Mechanism
two of the record's largest turns were arguable before any run. OQ-16's reconstruction target is the lm_head input, so the softmax-bottleneck identity settles it (F-23's dated note of 2026-10-18); and the first free baseline left token ids out while, at S3, the published analyses of MoE routing say routing follows token identity (F-24's dated note of the same day: consistent with the absorption measured, not a prediction of it, since the re-test added token ids and the standardized state in one arm). The registration apparatus checks procedure, not premises, so it could not catch either: a kill criterion and a pinned prediction test what was asked, not whether the question was already answered. sideband/docs/LESSONS.md:241 · 61a765b6939d
Measured cost
the OQ-16 lane through its Stage R run and Stage A freeze (F-22, F-23); the existence claims F-15 and F-17 and their re-tests F-24, F-25, F-28 and F-29, at readers below the practice the probing references describe (untuned, where Pimentel et al. ask for the highest-performing probe available; scored on top-1 accuracy; on raw inputs at F-15 and F-17). Counted in registered stages and findings, not hours: two stages of OQ-16 and four re-test findings. The cold read of 2026-10-18 (five reviewers briefed without the record's conclusions; docs/PUBLIC_CUT_PLAN.md 15.8) named both. [withheld: docs/PUBLIC_CUT_PLAN.md not in this repository; see WITHHELD.md] sideband/docs/LESSONS.md:249 · 61a765b6939d
Rule
before a freeze, write the strongest "already known, or settled by an identity" argument against the question; read the three nearest papers literally; read ten raw samples of the data in hand; run a tuned standard baseline. Formalise only what survives that page, and put the papers the argument needed into docs/PRIOR_ART.md before the registration, not after the result (its 2026-10-18 additions are the ones this record needed). sideband/docs/LESSONS.md:256 · 61a765b6939d
Not yet followed by
every registration in this record, all of which predate the rule, and the harness, which has no pre-registration checklist step for it. sideband/docs/LESSONS.md:262 · 61a765b6939d

Cites OQ-16, F-23, F-24, F-22, F-15, F-17, F-25, F-28, F-29

What the repository leaves out

The public repository was exported from private research repositories with a fresh history. Its 4 WITHHELD.md files, one at the root and one per folder, say what each folder leaves out (private planning documents, operator logs, session scratch) and what the export changed in place; here they are rendered from their lines. 2 lines are not shown because they quote a local path.

WITHHELD.md · 9 lines

What this repository leaves out WITHHELD.md:1 · 9efdd8169c92

This repository is a fresh-history export of private research repositories; each folder's WITHHELD.md lists what that folder leaves out and what the export changed in it: WITHHELD.md:3 · 9efdd8169c92

At the root, the export emits README.md, LICENSE, NOTICE, CITATION.cff, licenses/ and .github/workflows/ci.yml from authored sources kept in the private sideband repository under a withheld folder, and this file. The export tool itself and its test are withheld with the sideband folder (its WITHHELD.md says why). WITHHELD.md:9 · 9efdd8169c92

sideband/WITHHELD.md · 145 lines

What this folder leaves out, and what the export changed sideband/WITHHELD.md:1 · b01fa29dbd2b

Generated by the export tool of the private sideband repository at commit e3d7ea76b93dd5497aae99d75abc2fe590b7bce7 (2026-10-19). Every file in this folder that the export did not edit or generate is the tracked file at that commit, byte for byte (the export reads the index blobs and the public repository stores them unconverted), so a sha256 or git blob id the record cites for such a file reproduces here; a file the export edited (listed below: a withheld section, a replaced sentence, a filled placeholder or an appended marker) hashes differently, and a digest the commit log cites for one of those names the private bytes. The plan behind the cut is itself withheld; this file is its reader form. sideband/WITHHELD.md:3 · b01fa29dbd2b

The rest of this file: 8 more sections

Never tracked, so never here sideband/WITHHELD.md:5 · b01fa29dbd2b

  • harvest/: agent transcripts, downloaded papers and patents, the sibling manifests. sideband/WITHHELD.md:7 · b01fa29dbd2b
  • HANDOFF.md: the operator-local session log (docs/RELEASE_NOTES.md 3.3 and docs/GLOSSARY.md describe it). sideband/WITHHELD.md:8 · b01fa29dbd2b
  • cc/GOAL.md, cc/_working_summary.md, cc/blueprints/, cc/execution_plan.json: harness session state. sideband/WITHHELD.md:9 · b01fa29dbd2b
  • .claude/settings.json: per-machine hook wiring. sideband/WITHHELD.md:10 · b01fa29dbd2b
  • results/real/harvest_*/ arrays (*.npz) and results/real/*.byproduct/ folders: the 13.4 GB the sha256 bindings in the committed metas verify; the Zenodo deposit named in the root README carries them (the ignore file also names results/real/oq16_recon_weights.npz, a file never produced, docs/DATA.md; two small tracked weight files, k3_probe_weights.npz and oq15_head_weights.npz, DO ship, and NOTICE and docs/THIRD_PARTY.md's 2026-10-13 note describe them). sideband/WITHHELD.md:11 · b01fa29dbd2b

Tracked files withheld whole sideband/WITHHELD.md:13 · b01fa29dbd2b

PathWhySource line
docs/FUTURE_IDEAS.mdan idea list; its lockstep section carries commercial blueprints (plan section 3) (1 file)sideband/WITHHELD.md:17 · b01fa29dbd2b
docs/IDEAS.mdan idea list in form, nine ranked entries with verifier scores (plan section 3) (1 file)sideband/WITHHELD.md:18 · b01fa29dbd2b
docs/PUBLIC_CUT_PLAN.mdthe plan of this cut; this file is its reader form (plan section 3) (1 file)sideband/WITHHELD.md:19 · b01fa29dbd2b
docs/public_cut/the sources of the authored export-only additions; each is emitted to its public place and the folder is never shipped (plan 11.2) (80 files)sideband/WITHHELD.md:20 · b01fa29dbd2b
.github/the private tree's root workflow; the public repository's root workflow runs the same steps from inside this folder (plan 10.7) (1 file)sideband/WITHHELD.md:21 · b01fa29dbd2b
tools/export_public_cut.pythe export tool: its anchors quote the passages it withholds, so the tool cannot ship with them; this file describes what it does (plan 12) (1 file)sideband/WITHHELD.md:22 · b01fa29dbd2b
tests/test_export_public_cut.pythe export tool's test, withheld with the tool (plan 12) (1 file)sideband/WITHHELD.md:23 · b01fa29dbd2b
docs/audits/critic-recovered/2026-10-13-b10c3ea1/critic rounds of the public-cut lane (the export tool, commit A, 3e7b744): they review the withheld export tool, its test and the plan, quoting their text, digests and placeholders, and go with them (plan 12; 14.2 item 3) (10 files)sideband/WITHHELD.md:24 · b01fa29dbd2b
docs/audits/critic-recovered/2026-10-13-46731adb/critic rounds of the public-cut lane (commit B, the pinned export and its proof, 977ff11): they review the withheld export tool, its test and the plan, quoting their text, digests and placeholders, and go with them (plan 12; 14.2 item 3) (46 files)sideband/WITHHELD.md:25 · b01fa29dbd2b
docs/audits/critic-recovered/2026-10-14-a8adf32c/critic rounds of the public-cut lane (the front-doors lane, d481935): they review the withheld export tool, its test and the plan, quoting their text, digests and placeholders, and go with them (plan 12; 14.2 item 3) (73 files)sideband/WITHHELD.md:26 · b01fa29dbd2b
docs/audits/critic-recovered/2026-10-15-c8cb4ef4/critic rounds of the public-cut lane (decision batch lane 1, the gates, 51ceb9c): they review the withheld export tool, its test and the plan, quoting their text, digests and placeholders, and go with them (plan 12; 14.2 item 3) (1 file)sideband/WITHHELD.md:27 · b01fa29dbd2b
docs/audits/critic-recovered/2026-10-15-9dfaf64d/critic rounds of the public-cut lane (decision batch lane 2, the tool, the allowlist and the record, 5c264b0): they review the withheld export tool, its test and the plan, quoting their text, digests and placeholders, and go with them (plan 12; 14.2 item 3) (23 files)sideband/WITHHELD.md:28 · b01fa29dbd2b
docs/audits/critic-recovered/2026-10-03-58dc4aea/critic/REPORT.mdthe critic's copy of docs/REPORT.md as of 2026-10-03, carrying the placeholders the export fills in the tracked file (1 file)sideband/WITHHELD.md:29 · b01fa29dbd2b
docs/audits/critic-recovered/2026-10-02-12e6097a/strat/critic5/diff_ptrs.txta diff hunk of the withheld idea file read by the R1 landing's fifth round; its lines go with the file they come from (plan section 3) (1 file)sideband/WITHHELD.md:30 · b01fa29dbd2b
docs/audits/critic-recovered/MANIFEST_WITHHELD.jsonthe manifest rows of the withheld recovered files above, digests included; MANIFEST.json beside it lists them by session and count (plan 14.2 item 3) (1 file)sideband/WITHHELD.md:31 · b01fa29dbd2b

Tracked files changed in the export sideband/WITHHELD.md:33 · b01fa29dbd2b

Every change is marked in place with [withheld: ...; see WITHHELD.md], is a filled placeholder, or is a scrub marker (the Paths section below); the private file is untouched. sideband/WITHHELD.md:35 · b01fa29dbd2b

Lines that name a withheld file sideband/WITHHELD.md:70 · b01fa29dbd2b

Every line of a text file this export may edit that names one of the withheld files above carries the marker [withheld: <name> not in this repository; see WITHHELD.md], appended at export: inside the last cell of a table row, as a # comment in a code or config file. The names are the path (either separator, any leading directory), the basename, the stem where it carries an underscore, each in any case, and the forms listed as extra: (?-i:\bIDEAS\b)(?!\.md). The files the record freezes (results/, data/, tools/cc/, the paths .gitattributes marks -text or binary, and any frozen instruments the folder's pyproject.toml lists) are never edited; a mention there is listed under residue below. Exempt from the rule: .github/ (the public repository's root carries .github/workflows/ci.yml with the same steps plus the analysis extra, and the harness code under tools/cc/ names the path generically; every sentence found describing the private workflow's install contract carries an export note instead (each is listed under 'Tracked files changed in the export' as 'an export note on the private workflow's install line')); docs/audits/critic-recovered/2026-10-03-58dc4aea/critic/REPORT.md (the recovered critic folder's frozen README, manifests and scripts name this withheld file by design and take no marker, and its basename is a word of the record; WITHHELD.md lists what is withheld (plan 14.2 item 3)); docs/audits/critic-recovered/2026-10-02-12e6097a/strat/critic5/diff_ptrs.txt (the recovered critic folder's frozen README, manifests and scripts name this withheld file by design and take no marker, and its basename is a word of the record; WITHHELD.md lists what is withheld (plan 14.2 item 3)); docs/audits/critic-recovered/MANIFEST_WITHHELD.json (the recovered critic folder's frozen README, manifests and scripts name this withheld file by design and take no marker, and its basename is a word of the record; WITHHELD.md lists what is withheld (plan 14.2 item 3)). Lines marked, by file: sideband/WITHHELD.md:72 · b01fa29dbd2b

Provenance residue kept as is sideband/WITHHELD.md:102 · b01fa29dbd2b

one line is not shown here: it quotes a local path

  • docs/audits/critic-recovered/2026-10-02-12e6097a/strat/critic/block_now.txt: FUTURE_IDEAS: a recovered critic file quoting the record's own pointer to a withheld idea file; frozen bytes take no marker (plan 14.2 item 3). sideband/WITHHELD.md:112 · b01fa29dbd2b
  • docs/audits/critic-recovered/2026-10-02-12e6097a/strat/critic3/draft_norm.txt: FUTURE_IDEAS: a recovered critic file quoting the record's own pointer to a withheld idea file; frozen bytes take no marker (plan 14.2 item 3). sideband/WITHHELD.md:113 · b01fa29dbd2b
  • docs/audits/critic-recovered/2026-10-02-12e6097a/strat/critic3/wt_norm.txt: FUTURE_IDEAS: a recovered critic file quoting the record's own pointer to a withheld idea file; frozen bytes take no marker (plan 14.2 item 3). sideband/WITHHELD.md:114 · b01fa29dbd2b
  • docs/audits/critic-recovered/2026-10-02-12e6097a/strat/critic5/diff_bp.txt: FUTURE_IDEAS: a recovered critic file quoting the record's own pointer to a withheld idea file; frozen bytes take no marker (plan 14.2 item 3). sideband/WITHHELD.md:115 · b01fa29dbd2b
  • docs/audits/critic-recovered/2026-10-02-12e6097a/strat/critic5/outcome.txt: FUTURE_IDEAS: a recovered critic file quoting the record's own pointer to a withheld idea file; frozen bytes take no marker (plan 14.2 item 3). sideband/WITHHELD.md:116 · b01fa29dbd2b
  • docs/audits/critic-recovered/2026-10-02-12e6097a/strat/critic5/outcome_flat.txt: FUTURE_IDEAS: a recovered critic file quoting the record's own pointer to a withheld idea file; frozen bytes take no marker (plan 14.2 item 3). sideband/WITHHELD.md:117 · b01fa29dbd2b
  • docs/audits/critic-recovered/2026-10-02-12e6097a/strat/critic6/outcome.txt: FUTURE_IDEAS: a recovered critic file quoting the record's own pointer to a withheld idea file; frozen bytes take no marker (plan 14.2 item 3). sideband/WITHHELD.md:118 · b01fa29dbd2b
  • docs/audits/critic-recovered/2026-10-02-12e6097a/strat/critic6/outcome_n.txt: FUTURE_IDEAS: a recovered critic file quoting the record's own pointer to a withheld idea file; frozen bytes take no marker (plan 14.2 item 3). sideband/WITHHELD.md:119 · b01fa29dbd2b
  • docs/audits/critic-recovered/2026-10-02-12e6097a/strat/critic6/added.txt: FUTURE_IDEAS: the R1 landing diff the critic's sixth round read; its hunk of the withheld idea file is the dated status line the record also holds in OPEN_QUESTIONS.md and README.md; frozen bytes take no marker (plan 14.2 item 3). sideband/WITHHELD.md:120 · b01fa29dbd2b
  • docs/audits/critic-recovered/2026-10-01-64064aa4/critic/reflect_now.txt: IDEAS.md: a recovered critic file quoting the record's own pointer to a withheld idea file; frozen bytes take no marker (plan 14.2 item 3). sideband/WITHHELD.md:121 · b01fa29dbd2b
  • docs/audits/critic-recovered/2026-10-04-d880d7a7/critic/controls.py: IDEAS.md: a recovered critic file quoting the record's own pointer to a withheld idea file; frozen bytes take no marker (plan 14.2 item 3). sideband/WITHHELD.md:122 · b01fa29dbd2b
  • docs/audits/critic-recovered/2026-10-05-fdeffb05/critic/gen_probe.py: IDEAS.md: a recovered critic file quoting the record's own pointer to a withheld idea file; frozen bytes take no marker (plan 14.2 item 3). sideband/WITHHELD.md:123 · b01fa29dbd2b
  • docs/audits/critic-recovered/2026-10-05-fdeffb05/critic/gen_probe2.py: IDEAS.md: a recovered critic file quoting the record's own pointer to a withheld idea file; frozen bytes take no marker (plan 14.2 item 3). sideband/WITHHELD.md:124 · b01fa29dbd2b
  • ESPALIER_MEMORY.md: 07-review-and-ideas.md: a 2026-10-12 memory row naming three documents the lockstep folder withholds (lockstep/WITHHELD.md lists them); the memory file is a log, kept as written (plan 15.7). sideband/WITHHELD.md:125 · b01fa29dbd2b
  • CHANGELOG.md: Blueprints E: a line naming a proposal of a review the lockstep folder withholds (lockstep/WITHHELD.md lists it); kept as written, name-only provenance (plan section 3; 15.7). sideband/WITHHELD.md:126 · b01fa29dbd2b
  • docs/COMMIT_LOG.md: Blueprints E: a commit message of 2026-10-11, quoted by the generated commit log, naming proposals of a review the lockstep folder withholds (lockstep/WITHHELD.md lists it); name-only provenance (plan section 3; 15.7). sideband/WITHHELD.md:127 · b01fa29dbd2b
  • docs/ECOSYSTEM.md: Blueprint E: a line naming a proposal of a review the lockstep folder withholds (lockstep/WITHHELD.md lists it); kept as written, name-only provenance (plan section 3; 15.7). sideband/WITHHELD.md:128 · b01fa29dbd2b
  • docs/audits/critic-recovered/2026-10-16-4a635c95/critic/status_before.txt: test_export_public_cut: a recovered critic file holding a git status line that names the withheld test; frozen bytes take no marker (plan 14.2 item 3). sideband/WITHHELD.md:129 · b01fa29dbd2b

Record lines flagged for the operator sideband/WITHHELD.md:131 · b01fa29dbd2b

The export tool's deny list (the one applied to commit messages) is also run over ESPALIER_MEMORY.md, which the tone sweep excludes. Each hit blocks the export's verdict, and a line kept after review is recorded here by its reason (the matched text stays in the tool's spec): sideband/WITHHELD.md:133 · b01fa29dbd2b

  • kept: ESPALIER_MEMORY.md: the 2026-10-11 row's sentence restating plan section 2's bullet of the same date on what the public cut leaves out (the row's wording, not the bullet's). sideband/WITHHELD.md:135 · b01fa29dbd2b
  • kept: ESPALIER_MEMORY.md: the 2026-10-10 row's name for an unregistered prediction, the science sense of the word. sideband/WITHHELD.md:136 · b01fa29dbd2b
  • kept: ESPALIER_MEMORY.md: the 2026-10-11 row's verdict on lockstep's original idea: the clause names the framing the row then rejects (what stands, the row says, is the coder, the conformance vectors and the method record); kept by the operator's call PC-3 of 2026-10-14 (plan 13). sideband/WITHHELD.md:137 · b01fa29dbd2b

Paths sideband/WITHHELD.md:139 · b01fa29dbd2b

one line is not shown here: it quotes a local path

The commit log sideband/WITHHELD.md:143 · b01fa29dbd2b

docs/COMMIT_LOG.md is generated at export: the 186 private commit messages up to the export commit, newest first, with the paths above scrubbed. 8 messages name a private repository or match the export's deny list (monetary and release-posture wording; the regex is COMMIT_LOG_DENY in the withheld tool) and have their bodies replaced by a marker, their subjects keeping their shape with the match replaced by [private repository] or [withheld]. A log of fixed historical text, it sits outside the export's tone sweep (the exclusion list is the withheld tool's), and a marker or placeholder it quotes is not counted as one. sideband/WITHHELD.md:145 · b01fa29dbd2b

amber/WITHHELD.md · 74 lines

What this folder leaves out, and what the export changed amber/WITHHELD.md:1 · fd35447b9bd0

Generated by the export tool of the private amber repository at commit 0b061b54d92c0c6abe926dedab1bcd6394ee7c85 (2026-10-19). Every file in this folder that the export did not edit or generate is the tracked file at that commit, byte for byte (the export reads the index blobs and the public repository stores them unconverted), so a sha256 or git blob id the record cites for such a file reproduces here; a file the export edited (listed below: a withheld section, a replaced sentence, a filled placeholder or an appended marker) hashes differently, and a digest the commit log cites for one of those names the private bytes. The plan behind the cut is itself withheld; this file is its reader form. amber/WITHHELD.md:3 · fd35447b9bd0

The rest of this file: 8 more sections

Never tracked, so never here amber/WITHHELD.md:5 · fd35447b9bd0

  • results/p2/, results/p2_llama32/, results/p7/: the top-256 logit dumps (about 1 GB each; results/p2_dumps.sha256 and results/p2_tinyllama_dumps.sha256 pin them) and results/p7/, which the ignore file names though P7 never ran (F-14 row E14). amber/WITHHELD.md:7 · fd35447b9bd0

Tracked files withheld whole amber/WITHHELD.md:9 · fd35447b9bd0

PathWhySource line
.github/the private tree's workflow: it checks out the private organisation's two sibling repositories with a token secret; the public repository's root workflow runs amber's steps from inside this folder (plan 10.7, 15.2) (1 file)amber/WITHHELD.md:13 · fd35447b9bd0

Tracked files changed in the export amber/WITHHELD.md:15 · fd35447b9bd0

Every change is marked in place with [withheld: ...; see WITHHELD.md], is a filled placeholder, or is a scrub marker (the Paths section below); the private file is untouched. amber/WITHHELD.md:17 · fd35447b9bd0

Lines that name a withheld file amber/WITHHELD.md:30 · fd35447b9bd0

Every line of a text file this export may edit that names one of the withheld files above carries the marker [withheld: <name> not in this repository; see WITHHELD.md], appended at export: inside the last cell of a table row, as a # comment in a code or config file. The names are the path (either separator, any leading directory), the basename, the stem where it carries an underscore, each in any case. The files the record freezes (results/, data/, tools/cc/, the paths .gitattributes marks -text or binary, and any frozen instruments the folder's pyproject.toml lists) are never edited; a mention there is listed under residue below. Exempt from the rule: .github/ (the public repository's root carries .github/workflows/ci.yml with an amber job running the same steps from inside this folder, and no line of amber's prose names the private workflow (0 lines by git grep -i .github -- ':!.github' on 2026-10-17)). Lines marked, by file: amber/WITHHELD.md:32 · fd35447b9bd0

Provenance residue kept as is amber/WITHHELD.md:36 · fd35447b9bd0

  • results/errata_critic/c_e7.py: huggingface: a Hugging Face cache path under the author's home directory (a model snapshot's location, with the snapshot hash) in a frozen evidence file; the first name ships inside it as it does inside the record's own checkout paths (plan section 2; 15.2). amber/WITHHELD.md:38 · fd35447b9bd0
  • results/errata_dryruns/dry2.json: huggingface: a Hugging Face cache path under the author's home directory (a model snapshot's location, with the snapshot hash) in a frozen evidence file; the first name ships inside it as it does inside the record's own checkout paths (plan section 2; 15.2). amber/WITHHELD.md:39 · fd35447b9bd0
  • results/errata_dryruns/dry_b3.json: huggingface: a Hugging Face cache path under the author's home directory (a model snapshot's location, with the snapshot hash) in a frozen evidence file; the first name ships inside it as it does inside the record's own checkout paths (plan section 2; 15.2). amber/WITHHELD.md:40 · fd35447b9bd0
  • results/errata_dryruns/dry_b3b.json: huggingface: a Hugging Face cache path under the author's home directory (a model snapshot's location, with the snapshot hash) in a frozen evidence file; the first name ships inside it as it does inside the record's own checkout paths (plan section 2; 15.2). amber/WITHHELD.md:41 · fd35447b9bd0
  • results/errata_recompute.json: huggingface: a Hugging Face cache path under the author's home directory (a model snapshot's location, with the snapshot hash) in a frozen evidence file; the first name ships inside it as it does inside the record's own checkout paths (plan section 2; 15.2). amber/WITHHELD.md:42 · fd35447b9bd0
  • results/errata_recompute_3.json: huggingface: a Hugging Face cache path under the author's home directory (a model snapshot's location, with the snapshot hash) in a frozen evidence file; the first name ships inside it as it does inside the record's own checkout paths (plan section 2; 15.2). amber/WITHHELD.md:43 · fd35447b9bd0
  • results/p1_download.log: huggingface: a Hugging Face cache path under the author's home directory (a model snapshot's location, with the snapshot hash) in a frozen evidence file; the first name ships inside it as it does inside the record's own checkout paths (plan section 2; 15.2). amber/WITHHELD.md:44 · fd35447b9bd0
  • results/p2_llama32.log: huggingface: a Hugging Face cache path under the author's home directory (a model snapshot's location, with the snapshot hash) in a frozen evidence file; the first name ships inside it as it does inside the record's own checkout paths (plan section 2; 15.2). amber/WITHHELD.md:45 · fd35447b9bd0
  • results/errata_critic/c_e7.py: REPO = Path(: an absolute path into the author's own amber checkout at the top of a frozen critic recompute script (results/errata_critic/REPORT.md names the scripts); the path's parent folder word was the operator's call (15.5 item 1), decided keep on 2026-10-18 (15.8): see the decided lines below. amber/WITHHELD.md:46 · fd35447b9bd0
  • results/errata_critic/c_misc.py: REPO = Path(: an absolute path into the author's own amber checkout at the top of a frozen critic recompute script (results/errata_critic/REPORT.md names the scripts); the path's parent folder word was the operator's call (15.5 item 1), decided keep on 2026-10-18 (15.8): see the decided lines below. amber/WITHHELD.md:47 · fd35447b9bd0
  • results/errata_critic/c_p1.py: REPO = Path(: an absolute path into the author's own amber checkout at the top of a frozen critic recompute script (results/errata_critic/REPORT.md names the scripts); the path's parent folder word was the operator's call (15.5 item 1), decided keep on 2026-10-18 (15.8): see the decided lines below. amber/WITHHELD.md:48 · fd35447b9bd0
  • results/errata_critic/c_p2.py: REPO = Path(: an absolute path into the author's own amber checkout at the top of a frozen critic recompute script (results/errata_critic/REPORT.md names the scripts); the path's parent folder word was the operator's call (15.5 item 1), decided keep on 2026-10-18 (15.8): see the decided lines below. amber/WITHHELD.md:49 · fd35447b9bd0
  • results/errata_critic/c_e4grid.py: R=Path(: an absolute path into the author's own amber checkout at the top of a frozen critic recompute script (results/errata_critic/REPORT.md names the scripts); the path's parent folder word was the operator's call (15.5 item 1), decided keep on 2026-10-18 (15.8): see the decided lines below. amber/WITHHELD.md:50 · fd35447b9bd0
  • results/errata_critic/c_f10p4.py: R=Path(: an absolute path into the author's own amber checkout at the top of a frozen critic recompute script (results/errata_critic/REPORT.md names the scripts); the path's parent folder word was the operator's call (15.5 item 1), decided keep on 2026-10-18 (15.8): see the decided lines below. amber/WITHHELD.md:51 · fd35447b9bd0
  • results/errata_critic/c_order.py: D = Path(: an absolute path into the author's own amber checkout at the top of a frozen critic recompute script (results/errata_critic/REPORT.md names the scripts); the path's parent folder word was the operator's call (15.5 item 1), decided keep on 2026-10-18 (15.8): see the decided lines below. amber/WITHHELD.md:52 · fd35447b9bd0
  • src/amber/rankcensus.py: probelab: the name of a private sibling repository in the census instrument's module docstring, kept as name-only provenance (plan section 3) because two errata payloads pin this file's blob and a marker would move the bytes they cite (15.5). amber/WITHHELD.md:53 · fd35447b9bd0

Record lines flagged for the operator amber/WITHHELD.md:55 · fd35447b9bd0

Each hit blocks the export's verdict, and a line kept after review is recorded here by its reason (the matched text stays in the tool's spec): amber/WITHHELD.md:57 · fd35447b9bd0

  • kept: docs/ERRATA_PLAN.md: kept by the operator's call of 2026-10-18 (plan 15.9, decision 2; 15.5 item 2): the errata plan's step 1 of 2026-10-10, written before the public cut was decided, declared results/p1, p4, p5 and p6 (the Llama-3.2 generations of P1 as int32 token ids with the draft's rank shards, and the archives coded from them with their decode outputs under P4 to P6; 1,384 tracked files) private evidence not for release; they ship under the root Llama notice as sideband's own model-output files ship, a dated note under the bullet in amber's own record says so, and they are the evidence behind F-8 and F-11 to F-13, with every errata row loading P1. amber/WITHHELD.md:59 · fd35447b9bd0
  • kept: results/errata_critic/c_e7.py: kept by the operator's call of 2026-10-18 (plan 15.8, decision 1; 15.5 item 1): this critic recompute script of 2026-10-10 (results/errata_critic/REPORT.md names it) opens with an absolute path into the author's amber checkout, whose parent folder's name is a word of the deny list the export applies to commit messages; the file is frozen evidence and part of the critic's only recompute code for F-14's rows, so it ships whole and the word ships inside this one path line. amber/WITHHELD.md:60 · fd35447b9bd0
  • kept: results/errata_critic/c_misc.py: kept by the operator's call of 2026-10-18 (plan 15.8, decision 1; 15.5 item 1): this critic recompute script of 2026-10-10 (results/errata_critic/REPORT.md names it) opens with an absolute path into the author's amber checkout, whose parent folder's name is a word of the deny list the export applies to commit messages; the file is frozen evidence and part of the critic's only recompute code for F-14's rows, so it ships whole and the word ships inside this one path line. amber/WITHHELD.md:61 · fd35447b9bd0
  • kept: results/errata_critic/c_p1.py: kept by the operator's call of 2026-10-18 (plan 15.8, decision 1; 15.5 item 1): this critic recompute script of 2026-10-10 (results/errata_critic/REPORT.md names it) opens with an absolute path into the author's amber checkout, whose parent folder's name is a word of the deny list the export applies to commit messages; the file is frozen evidence and part of the critic's only recompute code for F-14's rows, so it ships whole and the word ships inside this one path line. amber/WITHHELD.md:62 · fd35447b9bd0
  • kept: results/errata_critic/c_p2.py: kept by the operator's call of 2026-10-18 (plan 15.8, decision 1; 15.5 item 1): this critic recompute script of 2026-10-10 (results/errata_critic/REPORT.md names it) opens with an absolute path into the author's amber checkout, whose parent folder's name is a word of the deny list the export applies to commit messages; the file is frozen evidence and part of the critic's only recompute code for F-14's rows, so it ships whole and the word ships inside this one path line. amber/WITHHELD.md:63 · fd35447b9bd0
  • kept: results/errata_critic/c_e4grid.py: kept by the operator's call of 2026-10-18 (plan 15.8, decision 1; 15.5 item 1): this critic recompute script of 2026-10-10 (results/errata_critic/REPORT.md names it) opens with an absolute path into the author's amber checkout, whose parent folder's name is a word of the deny list the export applies to commit messages; the file is frozen evidence and part of the critic's only recompute code for F-14's rows, so it ships whole and the word ships inside this one path line. amber/WITHHELD.md:64 · fd35447b9bd0
  • kept: results/errata_critic/c_f10p4.py: kept by the operator's call of 2026-10-18 (plan 15.8, decision 1; 15.5 item 1): this critic recompute script of 2026-10-10 (results/errata_critic/REPORT.md names it) opens with an absolute path into the author's amber checkout, whose parent folder's name is a word of the deny list the export applies to commit messages; the file is frozen evidence and part of the critic's only recompute code for F-14's rows, so it ships whole and the word ships inside this one path line. amber/WITHHELD.md:65 · fd35447b9bd0
  • kept: results/errata_critic/c_order.py: kept by the operator's call of 2026-10-18 (plan 15.8, decision 1; 15.5 item 1): this critic recompute script of 2026-10-10 (results/errata_critic/REPORT.md names it) opens with an absolute path into the author's amber checkout, whose parent folder's name is a word of the deny list the export applies to commit messages; the file is frozen evidence and part of the critic's only recompute code for F-14's rows, so it ships whole and the word ships inside this one path line. amber/WITHHELD.md:66 · fd35447b9bd0

Paths amber/WITHHELD.md:68 · fd35447b9bd0

Absolute paths into the author's own sideband checkouts stay as the record committed them (the export's own-checkout rule), as does a drive or root path with no user segment: the export's path rule concerns a user's home directory. docs/ERRATA_PLAN.md's two paths into that checkout (the harvest folder and the installed package) stay under the rule; its four paths into the amber and lockstep checkouts are scrubbed (the operator's call 1 of 2026-10-17); the paths into the amber checkout at the top of the frozen critic scripts under results/errata_critic/ are listed under residue above. In docs/ERRATA_PLAN.md, docs/COMMIT_LOG.md any other home-directory path is replaced by <path outside the repository, scrubbed at export>, and a digest (sha256 or blob id, full or 12-hex) of a withheld file by <digest of a withheld file, scrubbed at export>. amber/WITHHELD.md:70 · fd35447b9bd0

The commit log amber/WITHHELD.md:72 · fd35447b9bd0

docs/COMMIT_LOG.md is generated at export: the 67 private commit messages up to the export commit, newest first, with the paths above scrubbed. 0 messages name a private repository or match the export's deny list (monetary and release-posture wording; the regex is COMMIT_LOG_DENY in the withheld tool) and have their bodies replaced by a marker, their subjects keeping their shape with the match replaced by [private repository] or [withheld]. A log of fixed historical text, it sits outside the export's tone sweep (the exclusion list is the withheld tool's), and a marker or placeholder it quotes is not counted as one. amber/WITHHELD.md:74 · fd35447b9bd0

lockstep/WITHHELD.md · 62 lines

What this folder leaves out, and what the export changed lockstep/WITHHELD.md:1 · c57265724b55

Generated by the export tool of the private lockstep repository at commit a84bef4968dcd7effc1a619ae9fa1d36c8629c68 (2026-10-19). Every file in this folder that the export did not edit or generate is the tracked file at that commit, byte for byte (the export reads the index blobs and the public repository stores them unconverted), so a sha256 or git blob id the record cites for such a file reproduces here; a file the export edited (listed below: a withheld section, a replaced sentence, a filled placeholder or an appended marker) hashes differently, and a digest the commit log cites for one of those names the private bytes. The plan behind the cut is itself withheld; this file is its reader form. lockstep/WITHHELD.md:3 · c57265724b55

The rest of this file: 7 more sections

Never tracked, so never here lockstep/WITHHELD.md:5 · c57265724b55

Tracked files withheld whole lockstep/WITHHELD.md:11 · c57265724b55

PathWhySource line
docs/00-vision.mdthe planning document of the project's original framing: four of its five sections set out a commercial plan and a twelve-week target; its one-sentence version is quoted in README.md's first screen (plan 10.3, 15.6) (1 file)lockstep/WITHHELD.md:15 · c57265724b55
docs/07-review-and-ideas.mda review followed by six proposals for further work, two of them later built (the cost model and the kill-criteria document), one a separate venture; three (D, E and F) were carried into an idea list the sideband folder also withholds (plan 10.3, 15.6) (1 file)lockstep/WITHHELD.md:16 · c57265724b55
docs/03-patent-and-licensing-notes.mdengineering notes on patents and licensing written for a counsel review that never ran (TP-5, archived): unverified freedom-to-operate flags and a licence plan the release supersedes (plan 10.3; 15.1 call 2's shape, 15.6) (1 file)lockstep/WITHHELD.md:17 · c57265724b55
.github/the private tree's workflow (a Python 3.10 to 3.13 matrix of this folder's validation commands and a wheel smoke test); a workflow in a sub-folder never runs, and the public repository's root workflow runs the validation commands from inside this folder (plan 10.7, 15.6) (1 file)lockstep/WITHHELD.md:18 · c57265724b55

Tracked files changed in the export lockstep/WITHHELD.md:20 · c57265724b55

Every change is marked in place with [withheld: ...; see WITHHELD.md], is a filled placeholder, or is a scrub marker (the Paths section below); the private file is untouched. lockstep/WITHHELD.md:22 · c57265724b55

Lines that name a withheld file lockstep/WITHHELD.md:33 · c57265724b55

Every line of a text file this export may edit that names one of the withheld files above carries the marker [withheld: <name> not in this repository; see WITHHELD.md], appended at export: inside the last cell of a table row, as a # comment in a code or config file. The names are the path (either separator, any leading directory), the basename, the stem where it carries an underscore, each in any case, and the forms listed as extra: (?-i:\bdocs[\\/]0[037])(?![\w-]|\.\w); (?-i:\b0[037]-(?:vision|review-and-ideas|patent-and-licensing-notes))(?![\w-]|\.md); (?-i:\bBlueprints? [A-F](?:/[A-F])*)\b. The files the record freezes (results/, data/, tools/cc/, the paths .gitattributes marks -text or binary, and any frozen instruments the folder's pyproject.toml lists) are never edited; a mention there is listed under residue below. Exempt from the rule: .github/ (the public repository's root carries .github/workflows/ci.yml with a lockstep job running this folder's validation commands, and the harness code, its tests and its install manual name the path generically (tools/trellis_install_surface.py, tools/trellis_reflect.py, .claude/trellis-tests/test_public_project_surface.py, trellis-docs/trellis-harness.md)). Lines marked, by file: lockstep/WITHHELD.md:35 · c57265724b55

Provenance residue kept as is lockstep/WITHHELD.md:51 · c57265724b55

  • tools/lockstep_costmodel.py: PLACEHOLDER ratios/speeds: the cost model's own label for its assumed input constants, in the instrument behind F-4 and F-8, which prints the same warning after every scenario (plan 15.7). lockstep/WITHHELD.md:53 · c57265724b55
  • README.md: are carried into sideband's: the 2026-10-11 parking banner's pointer to an idea list the sideband folder withholds (sideband/WITHHELD.md lists it); the banner is kept as written beneath the first screen (plan 15.7). lockstep/WITHHELD.md:54 · c57265724b55

Paths lockstep/WITHHELD.md:56 · c57265724b55

Absolute paths into the author's own sideband checkouts stay as the record committed them (the export's own-checkout rule), as does a drive or root path with no user segment: the export's path rule concerns a user's home directory. task-packs/done/TP-7_errata_and_archive.md's two paths into the amber checkout are scrubbed; its path into the sideband checkout's harvest folder stays under the rule, as the sibling folders' paths into that checkout do. In task-packs/done/TP-7_errata_and_archive.md, docs/COMMIT_LOG.md any other home-directory path is replaced by <path outside the repository, scrubbed at export>, and a digest (sha256 or blob id, full or 12-hex) of a withheld file by <digest of a withheld file, scrubbed at export>. lockstep/WITHHELD.md:58 · c57265724b55

The commit log lockstep/WITHHELD.md:60 · c57265724b55

docs/COMMIT_LOG.md is generated at export: the 18 private commit messages up to the export commit, newest first, with the paths above scrubbed. 0 messages name a private repository or match the export's deny list (monetary and release-posture wording; the regex is COMMIT_LOG_DENY in the withheld tool) and have their bodies replaced by a marker, their subjects keeping their shape with the match replaced by [private repository] or [withheld]. A log of fixed historical text, it sits outside the export's tone sweep (the exclusion list is the withheld tool's), and a marker or placeholder it quotes is not counted as one. lockstep/WITHHELD.md:62 · c57265724b55

Modules: standalone copies of the record's tools

8 parts of the record's code and method that stand on their own, one catalogue card each, beside the record's notes on what it does, how to use it and what to know first. The packaged folders are not published as downloads; instead each card carries one command that builds its folder from the repository at tag v1.0.0, checking every file against its full sha256 before it writes any. The file list under each card links every file to the same file at that tag, its sha256 prefix checked against the tag when this page's data was extracted (65 files). The card text comes from the module catalogue kept with the private record, so it carries no source chips; a test line gives the counts from running that copy's tests.

exact verifier

Exact verifier: the speculative-decoding accept/reject step in numpy

sideband · folder

6 files, 29146 bytes; MIT licence.

Contents: spec_step; run_block; residual_distribution; analytic_acceptance; StepOutcome; BlockOutcome; RunMetrics.

Tests of this copy: 24 passed

How to use this tool

A speculative decoder lets a small draft model guess the next tokens and a large model check them. This is the checking rule: it keeps or replaces each guess so that the text that comes out follows the large model's own probabilities exactly, whatever the draft guessed. The tests prove that by counting through every outcome.

  1. Copy the command under Get this folder, run it, then run the tests to see the rule hold.
  2. In your code, import spec_step from sideband.specdec. Give it a numpy random generator, then the large model's probabilities, then the draft's, for one position over the same vocabulary.
  3. It returns a StepOutcome: the token the draft proposed, whether it was kept, and the token that came out.
  4. For a block of several guesses use run_block, with one more large model distribution than draft distributions; the extra one gives a bonus token when every guess is kept.
  5. If you write a faster version, make these tests pass against it before you trust it.

A numpy implementation of the exact speculative-decoding accept/reject step (Leviathan et al. 2023) with tests that check by enumeration that the emitted token follows the target distribution for any draft distribution; for researchers building or changing speculative decoders.

Get this folder

One command makes a new folder named exact-verifier inside the folder your terminal is in. It changes nothing else, joins no project and installs nothing. Paste it into bash, zsh or PowerShell with Python on the path; where python is not found, change the first word to python3. It downloads the 6 files listed below from the repository at tag v1.0.0, checks each against its full sha256, writes nothing unless every one matches and the folder is new, and adds the pytest.ini the tests run under. The tests need numpy and pytest: pip install numpy pytest, then cd exact-verifier and python -m pytest.

Show the command
python -c "
import hashlib, pathlib, urllib.request
base = 'https://raw.githubusercontent.com/Mike-Byrne-AI/specdec-sidechannels/v1.0.0/'
out = pathlib.Path('exact-verifier')
files = [
    ('LICENSE', 'LICENSE', '8497cbc80c24245fc8f9be355fca4a8d15b0235b7de4870b30d573189092aeef'),
    ('src/sideband/__init__.py', 'sideband/src/sideband/__init__.py', 'b3123882e5c63944e9ecadcf31126ddbd513be73b9829ae89ee095dca685946f'),
    ('src/sideband/dists.py', 'sideband/src/sideband/dists.py', 'a8fb156eff2141061fa7b7d1ae61de364797aebb77acf7a7425f25fb7423e280'),
    ('src/sideband/specdec.py', 'sideband/src/sideband/specdec.py', '80502d0f5695a56f650aa1d166afaa7b34683628480bed703b8fd34f4640c65d'),
    ('src/sideband/steering.py', 'sideband/src/sideband/steering.py', '3271311b2303ba317036f01ca581a16da00bb22598eefd65e65d50a6b55ffff7'),
    ('tests/test_specdec.py', 'sideband/tests/test_specdec.py', '4791d46cc47cdf1c3d55623e7d6b203908375fff7fbdc136112c0348aff63ca2'),
]
if out.exists():
    raise SystemExit(str(out) + ' already exists: nothing written')
got = []
for dest, src, sha in files:
    data = urllib.request.urlopen(base + src).read()
    if hashlib.sha256(data).hexdigest() == sha:
        got.append((out / dest, data))
    else:
        raise SystemExit(src + ': its sha256 is not the one on the page: nothing written')
for path, data in got:
    path.parent.mkdir(parents=True, exist_ok=True)
    path.write_bytes(data)
(out / 'pytest.ini').write_bytes(chr(10).join(['[pytest]', 'testpaths = tests', 'pythonpath = src', '']).encode())
print('checked and wrote', len(got), 'files to', out.resolve())
"

In the record

In the record this file is the single implementation of accept, reject and residual (DECISIONS.md D-3: the real-model adapter converts logits to numpy and calls spec_step, and anyone re-implementing the verifier for speed must first make these numpy tests pass against their implementation). Its docstring states the property the tests check by enumeration: for any draft distribution, steered or not, the emitted token is distributed exactly as the target distribution, so steering can only change speed, never output quality.

Use

  • Copy src/sideband/ into your project, or vendor specdec.py and dists.py together (specdec imports normalize, sample and tv_distance from dists).
  • steering.py is here only because the test file uses its bucket_bias; it is not needed to call spec_step.
  • Run python -m pytest from this folder: pytest.ini points it at the copied tests and this folder's src/.

Good to know

  • Per-step acceptance is analytically 1 - TV(p, q), and that identity is tested too.
  • The docstring names the record's real-model adapter (sideband.real.hf_specdec, not copied) and the test file, which is copied here.
The 6 files, each linked to the same file at v1.0.0
Module pathPublic file at v1.0.0sha256Bytes
LICENSELICENSE8497cbc80c241070
src/sideband/__init__.pysideband/src/sideband/__init__.pyb3123882e5c687
src/sideband/dists.pysideband/src/sideband/dists.pya8fb156eff212115
src/sideband/specdec.pysideband/src/sideband/specdec.py80502d0f56957210
src/sideband/steering.pysideband/src/sideband/steering.py3271311b23036217
tests/test_specdec.pysideband/tests/test_specdec.py4791d46cc47c12447

cluster bootstrap stats

Cluster bootstrap: confidence intervals for prompt-clustered evaluations

sideband · folder

7 files, 35410 bytes; MIT licence.

Contents: clopper_pearson; cluster_bootstrap_ci; paired_cluster_delta_ci; paired_ratio_delta_ci; paired_ratio_delta_of_deltas_ci; e6_gate.pair; e6_gate.evaluate.

Tests of this copy: 38 passed, 2 skipped

How to use this tool

When two methods are scored on the same prompts, results from one prompt tend to move together, so treating every result as independent makes the uncertainty look smaller than it is. These functions give confidence intervals that resample whole prompts, and when comparing two methods they resample both with the same draw, so the difference is a fair paired one.

  1. Copy the command under Get this folder, run it, then run the tests.
  2. For each prompt, count the successes of method A, the successes of method B and the number of trials, with the prompts in the same order in all three arrays.
  3. Call paired_cluster_delta_ci with those three arrays. It returns the difference between the two success rates and the low and high ends of its 95% interval.
  4. For one method alone use cluster_bootstrap_ci. With nothing to group by, clopper_pearson gives the exact interval for k successes in n trials.
  5. The same seed always gives the same interval, so fix the seed before you look at results.

Numpy-only confidence intervals for clustered evaluation data: an exact Clopper-Pearson interval, a prompt-cluster percentile bootstrap, and paired delta and paired ratio-delta bootstraps that apply one resampling draw to both arms; for researchers comparing two methods on the same prompts.

Get this folder

One command makes a new folder named cluster-bootstrap-stats inside the folder your terminal is in. It changes nothing else, joins no project and installs nothing. Paste it into bash, zsh or PowerShell with Python on the path; where python is not found, change the first word to python3. It downloads the 7 files listed below from the repository at tag v1.0.0, checks each against its full sha256, writes nothing unless every one matches and the folder is new, and adds the pytest.ini the tests run under. The tests need numpy and pytest: pip install numpy pytest, then cd cluster-bootstrap-stats and python -m pytest.

Show the command
python -c "
import hashlib, pathlib, urllib.request
base = 'https://raw.githubusercontent.com/Mike-Byrne-AI/specdec-sidechannels/v1.0.0/'
out = pathlib.Path('cluster-bootstrap-stats')
files = [
    ('LICENSE', 'LICENSE', '8497cbc80c24245fc8f9be355fca4a8d15b0235b7de4870b30d573189092aeef'),
    ('src/sideband/__init__.py', 'sideband/src/sideband/__init__.py', 'b3123882e5c63944e9ecadcf31126ddbd513be73b9829ae89ee095dca685946f'),
    ('src/sideband/dists.py', 'sideband/src/sideband/dists.py', 'a8fb156eff2141061fa7b7d1ae61de364797aebb77acf7a7425f25fb7423e280'),
    ('src/sideband/stats.py', 'sideband/src/sideband/stats.py', '2def9e60b56a555740a98015ef44a73866712de002787fa277ac2042a880d4cd'),
    ('tests/test_e6_gate.py', 'sideband/tests/test_e6_gate.py', 'ce11bb5c19a06e91bee6d281b3f2db0cb4f16a75bc3db3d7e06c1749999e3a20'),
    ('tests/test_stats.py', 'sideband/tests/test_stats.py', 'a8b71dd3b498b35116235da824df6f79e4fd885af1e30fb7e858bdead2e9613c'),
    ('tools/e6_gate.py', 'sideband/tools/e6_gate.py', 'ca1f5e7ae7ce77f8fb44268098a8a762cf9599d5b56f691ae6703db5e93f79d5'),
]
if out.exists():
    raise SystemExit(str(out) + ' already exists: nothing written')
got = []
for dest, src, sha in files:
    data = urllib.request.urlopen(base + src).read()
    if hashlib.sha256(data).hexdigest() == sha:
        got.append((out / dest, data))
    else:
        raise SystemExit(src + ': its sha256 is not the one on the page: nothing written')
for path, data in got:
    path.parent.mkdir(parents=True, exist_ok=True)
    path.write_bytes(data)
(out / 'pytest.ini').write_bytes(chr(10).join(['[pytest]', 'testpaths = tests', 'pythonpath = src', '']).encode())
print('checked and wrote', len(got), 'files to', out.resolve())
"

In the record

stats.py is the estimator the record's root README recomputes a finding with in its worked example (paired_ratio_delta_ci over a committed state file), and the one its conversion gate, tools/e6_gate.py, imports; several record instruments carry bootstraps of their own beside it. Its docstring reports the pair of estimators together on purpose: the exact binomial interval assumes independent trials, which positions within a prompt are not, and the percentile bootstrap under-covers at small cluster counts (its docstring gives the measured coverage at 10, 30 and 60 clusters for a nominal 0.95), so neither end is conservative and both are reported.

Use

  • stats.py needs only the Array alias from dists.py; copy both, or inline the alias.
  • tools/e6_gate.py is a worked example of the paired-delta gate, not an API: its population table, results path and bar constants are the record's own runs.
  • Run python -m pytest from this folder. Two legs of test_e6_gate skip here because they read the record's committed verdict summaries.

Good to know

  • The Clopper-Pearson reference values in the tests were generated once from scipy.stats.beta and committed; scipy is not a dependency.
  • The bootstrap is deterministic given its seed; the paired forms apply one resampling draw to both arms, which is what makes the delta a paired delta.
The 7 files, each linked to the same file at v1.0.0
Module pathPublic file at v1.0.0sha256Bytes
LICENSELICENSE8497cbc80c241070
src/sideband/__init__.pysideband/src/sideband/__init__.pyb3123882e5c687
src/sideband/dists.pysideband/src/sideband/dists.pya8fb156eff212115
src/sideband/stats.pysideband/src/sideband/stats.py2def9e60b56a11457
tests/test_e6_gate.pysideband/tests/test_e6_gate.pyce11bb5c19a04057
tests/test_stats.pysideband/tests/test_stats.pya8b71dd3b4988877
tools/e6_gate.pysideband/tools/e6_gate.pyca1f5e7ae7ce7747

rank census

Rank census: does a model's token ranking reproduce across environments?

amber · folder

4 files, 17769 bytes; MIT licence.

Contents: rank_census; rank_census_topk; calibrate_tau; calibrate_tau_topk; RankCensus; stable_topk; top2_gap; canonical_topk; aligned_delta.

Tests of this copy: 10 passed

How to use this tool

The same model on two machines, or at two batch sizes, can rank its candidate next tokens a little differently. This tool takes the scores from both runs at the same positions and reports how often the top token, the top few tokens and their order change, and how close the top two scores were where the top token changed.

  1. Copy the command under Get this folder, run it, then run the tests.
  2. Save the model's scores (logits) at the same positions from each run as a numpy array with one row per position and one column per vocabulary entry.
  3. Call rank_census with the reference run first and the other run second.
  4. Read the RankCensus it returns: how often the top token flipped, how often the top 1, 8 and 64 changed as a set and in order (the ks argument sets those sizes), and the smallest, median and largest gap between the top two scores where the top token flipped.
  5. Its only dependency is numpy, so you can copy rankcensus.py on its own.

A numpy instrument that compares two logit dumps of the same positions (for example two devices or batch sizes) and reports top-1, top-k set and top-k order flip rates, the top-2 gap at each flip, a calibrated tolerance and flip burstiness; for researchers checking whether a model's token ranking reproduces across environments.

Get this folder

One command makes a new folder named rank-census inside the folder your terminal is in. It changes nothing else, joins no project and installs nothing. Paste it into bash, zsh or PowerShell with Python on the path; where python is not found, change the first word to python3. It downloads the 4 files listed below from the repository at tag v1.0.0, checks each against its full sha256, writes nothing unless every one matches and the folder is new, and adds the pytest.ini the tests run under. The tests need numpy and pytest: pip install numpy pytest, then cd rank-census and python -m pytest.

Show the command
python -c "
import hashlib, pathlib, urllib.request
base = 'https://raw.githubusercontent.com/Mike-Byrne-AI/specdec-sidechannels/v1.0.0/'
out = pathlib.Path('rank-census')
files = [
    ('LICENSE', 'LICENSE', '8497cbc80c24245fc8f9be355fca4a8d15b0235b7de4870b30d573189092aeef'),
    ('src/amber/__init__.py', 'amber/src/amber/__init__.py', 'e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855'),
    ('src/amber/rankcensus.py', 'amber/src/amber/rankcensus.py', 'ede73b25d417ec7187c7bedb0a556587ffbba333825f61bf9d0deb745638724c'),
    ('tests/test_rankcensus.py', 'amber/tests/test_rankcensus.py', '8693ab4f8634e297cbe614bce286210d31d7327ab4a3df10b5f740a1aeee14cb'),
]
if out.exists():
    raise SystemExit(str(out) + ' already exists: nothing written')
got = []
for dest, src, sha in files:
    data = urllib.request.urlopen(base + src).read()
    if hashlib.sha256(data).hexdigest() == sha:
        got.append((out / dest, data))
    else:
        raise SystemExit(src + ': its sha256 is not the one on the page: nothing written')
for path, data in got:
    path.parent.mkdir(parents=True, exist_ok=True)
    path.write_bytes(data)
(out / 'pytest.ini').write_bytes(chr(10).join(['[pytest]', 'testpaths = tests', 'pythonpath = src', '']).encode())
print('checked and wrote', len(got), 'files to', out.resolve())
"

In the record

In amber this is the instrument for its rank-stability question: amber's determinism contract is that the draft's token ordering reproduces at decode time, not its logits bitwise, and this module measures how far two logit dumps of the same teacher-forced positions disagree in exactly that quantity. Its docstring explains why every top-1 flip carries the reference top-2 gap (a flip at a hairline gap is floating-point noise, a flip at a wide gap is a bug) and why the tolerance is calibrated from a benign perturbation rather than picked.

Use

  • One file with no internal imports: copy src/amber/rankcensus.py alone, or the folder.
  • Run python -m pytest from this folder.

Good to know

  • The docstring credits the owner's other research projects by name as the sources of its rules (the calibrated tolerance, the near-tie budget, the drift-census shape); those are provenance notes, not imports, and would be reworded for a standalone card.
  • Order convention is descending value, ascending index on ties, the same as amber's codec, so the instrument and the codec agree on what a rank is.
The 4 files, each linked to the same file at v1.0.0
Module pathPublic file at v1.0.0sha256Bytes
LICENSELICENSE8497cbc80c241070
src/amber/__init__.pyamber/src/amber/__init__.pye3b0c44298fc0
src/amber/rankcensus.pyamber/src/amber/rankcensus.pyede73b25d41711471
tests/test_rankcensus.pyamber/tests/test_rankcensus.py8693ab4f86345228

synthetic testbed

Synthetic testbed: a CPU world for draft-steering experiments

sideband · folder

14 files, 78787 bytes; MIT licence.

Contents: WorldConfig; World; TelemetryDataset; run_simulation; train_sidecar; SimResult; POLICIES; HashSketch; SoftmaxRegression; GatedSteerer; posterior_bias; bucket_bias.

Tests of this copy: 68 passed

How to use this tool

A small made up world that runs on an ordinary CPU: a large model, a draft model and a stream of side information, all generated from a hidden topic that changes now and then. Because the hidden truth is known, you can try an idea for steering the draft, and check the code behind it, before spending time on real models.

  1. Copy the command under Get this folder, run it, then run the tests.
  2. Make a world with World(WorldConfig()); change the fields of WorldConfig, its seed among them, to shape it.
  3. Call run_simulation(world, seed, n_positions, gamma, policy=...) once for each name in POLICIES and compare the SimResult each returns; gamma is the number of draft guesses per check.
  4. The learned policy needs a trained model first, which train_sidecar makes.
  5. Know its limit: here the side information and the large model come from the same hidden topic, so no steering can beat the free baseline in this world. Whether real models allow more is a question it cannot answer.

A numpy synthetic world (a sticky concept chain with target, draft and hashed telemetry) wired to an exact speculative decoder and five policies (vanilla, two oracles, the free-information baseline and a learned sidecar), so steering and analysis code can be checked against known ground truth on CPU; for researchers prototyping draft-steering ideas.

Get this folder

One command makes a new folder named synthetic-testbed inside the folder your terminal is in. It changes nothing else, joins no project and installs nothing. Paste it into bash, zsh or PowerShell with Python on the path; where python is not found, change the first word to python3. It downloads the 14 files listed below from the repository at tag v1.0.0, checks each against its full sha256, writes nothing unless every one matches and the folder is new, and adds the pytest.ini the tests run under. The tests need numpy and pytest: pip install numpy pytest, then cd synthetic-testbed and python -m pytest.

Show the command
python -c "
import hashlib, pathlib, urllib.request
base = 'https://raw.githubusercontent.com/Mike-Byrne-AI/specdec-sidechannels/v1.0.0/'
out = pathlib.Path('synthetic-testbed')
files = [
    ('LICENSE', 'LICENSE', '8497cbc80c24245fc8f9be355fca4a8d15b0235b7de4870b30d573189092aeef'),
    ('src/sideband/__init__.py', 'sideband/src/sideband/__init__.py', 'b3123882e5c63944e9ecadcf31126ddbd513be73b9829ae89ee095dca685946f'),
    ('src/sideband/buckets.py', 'sideband/src/sideband/buckets.py', '6a67905cfed5cd15dc0caffbfd8953f5d206cdc4b3291e6093bbbf00bf1b60b9'),
    ('src/sideband/dists.py', 'sideband/src/sideband/dists.py', 'a8fb156eff2141061fa7b7d1ae61de364797aebb77acf7a7425f25fb7423e280'),
    ('src/sideband/pipeline.py', 'sideband/src/sideband/pipeline.py', '8262c73cbcf67105aa1ea416f453faa09bf7d34c4fc75b27fb365a348ad5115b'),
    ('src/sideband/sidecar.py', 'sideband/src/sideband/sidecar.py', '50b5c5ed452759365def8084d2cd36629d60f3b19d0c3a649de68e8f133363af'),
    ('src/sideband/sketch.py', 'sideband/src/sideband/sketch.py', '276b46e4ee26e175c7becc6bd7b3721e409b19392d9ed6b55c7a64c8ec228295'),
    ('src/sideband/specdec.py', 'sideband/src/sideband/specdec.py', '80502d0f5695a56f650aa1d166afaa7b34683628480bed703b8fd34f4640c65d'),
    ('src/sideband/steering.py', 'sideband/src/sideband/steering.py', '3271311b2303ba317036f01ca581a16da00bb22598eefd65e65d50a6b55ffff7'),
    ('src/sideband/synthetic.py', 'sideband/src/sideband/synthetic.py', '388b184ccbf08ac3cad7818dc309d546c3b9d3bf41e87c074140b0ff6561f148'),
    ('tests/test_posterior_mechanism.py', 'sideband/tests/test_posterior_mechanism.py', '5a3c10c736b9325e81c476fbd6db15d973657f7c334573c65abfd80161ea6ddd'),
    ('tests/test_sketch_buckets.py', 'sideband/tests/test_sketch_buckets.py', '5e7996566a7c3252bcd474090f5b9ee11b00cac7f5ecbbf8667fff300db3d5b2'),
    ('tests/test_steering_pipeline.py', 'sideband/tests/test_steering_pipeline.py', '09d16f839058cb537e62dd4f882fbccd261cdb547af9cbfba3c0880e22effc70'),
    ('tests/test_synthetic_sidecar.py', 'sideband/tests/test_synthetic_sidecar.py', '215b0100cd0e5f843ba7d37781a80b6ae5ce8bb7bf01c4ec61e9657a20aa526e'),
]
if out.exists():
    raise SystemExit(str(out) + ' already exists: nothing written')
got = []
for dest, src, sha in files:
    data = urllib.request.urlopen(base + src).read()
    if hashlib.sha256(data).hexdigest() == sha:
        got.append((out / dest, data))
    else:
        raise SystemExit(src + ': its sha256 is not the one on the page: nothing written')
for path, data in got:
    path.parent.mkdir(parents=True, exist_ok=True)
    path.write_bytes(data)
(out / 'pytest.ini').write_bytes(chr(10).join(['[pytest]', 'testpaths = tests', 'pythonpath = src', '']).encode())
print('checked and wrote', len(got), 'files to', out.resolve())
"

In the record

The record's findings index says its first arc built the verifier and this testbed and used it to fix the shapes that every later experiment inherits (docs/FINDINGS_INDEX.md, the Synthetic world arc), and two of its decisions live in pipeline.py: the free-information baseline policy (DECISIONS.md D-5: any telemetry-steering claim is compared against steering from the target's own distribution at the last verified position, which every speculative decoder computes anyway) and the isolated telemetry random stream (D-6: policies are compared at identical seeds so trajectories are paired, and telemetry draws from its own generator so a policy that reads it cannot perturb the shared draft and verify trajectory).

Use

  • Copy src/sideband/ whole: pipeline.py imports synthetic, sidecar, steering, specdec and dists, and synthetic.py imports sketch.
  • buckets.py is here to keep test_sketch_buckets.py whole; synthetic.py does not import it.
  • Run python -m pytest from this folder.

Good to know

  • Regime warning from synthetic.py's own docstring: telemetry and the target's next-token distribution are functions of the same latent concept here, so a telemetry sidecar can at best match the free-information baseline, and whether real hidden-state sketches carry more is a question this world cannot answer by construction.
  • The copied files cite the record's decision, finding and open-question ids (D-3, D-5, D-6, D-7, D-8, D-9, D-10, D-11, F-3, F-7, F-8, OQ-1, OQ-4); those read against the record's DECISIONS.md, FINDINGS.md and OPEN_QUESTIONS.md.
The 14 files, each linked to the same file at v1.0.0
Module pathPublic file at v1.0.0sha256Bytes
LICENSELICENSE8497cbc80c241070
src/sideband/__init__.pysideband/src/sideband/__init__.pyb3123882e5c687
src/sideband/buckets.pysideband/src/sideband/buckets.py6a67905cfed54702
src/sideband/dists.pysideband/src/sideband/dists.pya8fb156eff212115
src/sideband/pipeline.pysideband/src/sideband/pipeline.py8262c73cbcf610907
src/sideband/sidecar.pysideband/src/sideband/sidecar.py50b5c5ed45273476
src/sideband/sketch.pysideband/src/sideband/sketch.py276b46e4ee262435
src/sideband/specdec.pysideband/src/sideband/specdec.py80502d0f56957210
src/sideband/steering.pysideband/src/sideband/steering.py3271311b23036217
src/sideband/synthetic.pysideband/src/sideband/synthetic.py388b184ccbf08735
tests/test_posterior_mechanism.pysideband/tests/test_posterior_mechanism.py5a3c10c736b95825
tests/test_sketch_buckets.pysideband/tests/test_sketch_buckets.py5e7996566a7c5605
tests/test_steering_pipeline.pysideband/tests/test_steering_pipeline.py09d16f83905815594
tests/test_synthetic_sidecar.pysideband/tests/test_synthetic_sidecar.py215b0100cd0e4809

findings lint

Findings lint: no regime line and no tracked source, no finding

sideband · folder

3 files, 17354 bytes; MIT licence.

Contents: main; lint_text; parse_entries; source_paths; source_lines.

Tests of this copy: 1 failed, 4 passed

How to use this tool

A check for a research log. Every finding in FINDINGS.md must say under what conditions it was measured (a regime line) and point to evidence files that git tracks (source lines). The check fails when a finding is missing either, or cites a file that does not exist or that git does not track.

  1. Copy the command under Get this folder and run it, then copy lint_findings.py from the new folder's tools folder into a tools folder at the root of your repository. It finds the root from where it sits and reads FINDINGS.md there.
  2. Give each finding a heading that starts ## F-, a line that starts - regime: and one or more lines that start - source: naming files under results/, tests/, docs/ or the other folders listed at the top of the file.
  3. Run python tools/lint_findings.py. It prints FAIL and exits with code 1 when a finding fails; add it to your make lint target and to your CI so it runs every time.
  4. Change the list of evidence folders at the top of the file to fit your layout.
  5. In this folder one of its five tests fails as expected: it reads the record's own FINDINGS.md, which is not copied.

A stdlib-only lint that fails a FINDINGS.md whose entries lack a regime line or a source line pointing at paths git tracks; for research repositories kept by people or coding agents, wired to make lint and CI.

Get this folder

One command makes a new folder named findings-lint inside the folder your terminal is in. It changes nothing else, joins no project and installs nothing. Paste it into bash, zsh or PowerShell with Python on the path; where python is not found, change the first word to python3. It downloads the 3 files listed below from the repository at tag v1.0.0, checks each against its full sha256, writes nothing unless every one matches and the folder is new, and adds the pytest.ini the tests run under. The tests need pytest: pip install pytest, then cd findings-lint and python -m pytest.

Show the command
python -c "
import hashlib, pathlib, urllib.request
base = 'https://raw.githubusercontent.com/Mike-Byrne-AI/specdec-sidechannels/v1.0.0/'
out = pathlib.Path('findings-lint')
files = [
    ('LICENSE', 'LICENSE', '8497cbc80c24245fc8f9be355fca4a8d15b0235b7de4870b30d573189092aeef'),
    ('tests/test_lint_findings.py', 'sideband/tests/test_lint_findings.py', '797bab9c41eb1c9fb6cdf08f0318f627b271b7f987213957bb25aa70c458258f'),
    ('tools/lint_findings.py', 'sideband/tools/lint_findings.py', '5ed53a4ca06a58200ed60a2e20fd55775afa042a890e27e25bd0b1acc1bb622e'),
]
if out.exists():
    raise SystemExit(str(out) + ' already exists: nothing written')
got = []
for dest, src, sha in files:
    data = urllib.request.urlopen(base + src).read()
    if hashlib.sha256(data).hexdigest() == sha:
        got.append((out / dest, data))
    else:
        raise SystemExit(src + ': its sha256 is not the one on the page: nothing written')
for path, data in got:
    path.parent.mkdir(parents=True, exist_ok=True)
    path.write_bytes(data)
(out / 'pytest.ini').write_bytes(chr(10).join(['[pytest]', 'testpaths = tests', '']).encode())
print('checked and wrote', len(got), 'files to', out.resolve())
"

In the record

The record's decision D-4 (findings carry regime and evidence, lint-enforced; its CLAUDE.md puts it as 'no regime, no finding') says every FINDINGS entry must have a regime line and a source line with existing evidence paths, and this lint is what enforces it on every make lint and CI run. METHOD.md's section on reusing the method in another repository names this file first of its three-file kit.

Use

  • Copy tools/lint_findings.py into a tools/ folder under your repository root: it finds the root from its own location and reads FINDINGS.md there. Run python tools/lint_findings.py; wire it to make lint and CI.
  • The prefixes it checks source paths under (results/, experiments/, tests/, docs/, src/, data/, tools/) are a constant at the top of the file; change them for your layout.
  • Run python -m pytest from this folder: one of the five tests reads the record's own FINDINGS.md, which is not in this folder, and fails with a file-not-found error; the other four pass on fixtures.

Good to know

  • amber carries a different, shorter tool under the same file name; this is sideband's.
  • The lint checks that cited paths exist and that git tracks them; when git cannot answer it falls back to the existence check and says so in its OK line, except under CI (the CI environment variable), where that is a failure.
The 3 files, each linked to the same file at v1.0.0
Module pathPublic file at v1.0.0sha256Bytes
LICENSELICENSE8497cbc80c241070
tests/test_lint_findings.pysideband/tests/test_lint_findings.py797bab9c41eb5499
tools/lint_findings.pysideband/tools/lint_findings.py5ed53a4ca06a10785

landing text

Landing text: generated prose whose every number traces to a payload field

sideband · folder

3 files, 52561 bytes; MIT licence.

Contents: Pair; Region; Doc; fill_template; score; assert_banned; unknown_numbers; check_texts; main.

Tests of this copy: 22 passed, 1 skipped

How to use this tool

A helper for the text that reports a result. Instead of typing numbers into a sentence, you pass each one through the tool, which records where it came from; a check then flags any number in the finished text that no result file holds. It also blocks banned phrases and reports mixed line endings.

  1. Copy the command under Get this folder, run it, then run the tests.
  2. In the script that writes your text, load your primary and replication result files with Pair and insert every number through it. Pair expects this record's result layout, so adapt its field names to yours first.
  3. Call Pair.write with a folder and your texts; it saves them beside a numbers.json file listing every number used.
  4. Run python tools/landing_text.py check numbers.json --files your_text.md. Each number it cannot trace is printed as a FINDING and the exit code is 1.
  5. python tools/landing_text.py eol reports the line ending of each file.

A stdlib-only tool for writing the text that lands a result: pairs of generated sentences and the payload fields they quote, a check that no number in the text is unknown to a payload, a banned-phrase assertion, a line-ending normaliser and a CLI; for research repositories whose results are landed by a coding agent.

Get this folder

One command makes a new folder named landing-text inside the folder your terminal is in. It changes nothing else, joins no project and installs nothing. Paste it into bash, zsh or PowerShell with Python on the path; where python is not found, change the first word to python3. It downloads the 3 files listed below from the repository at tag v1.0.0, checks each against its full sha256, writes nothing unless every one matches and the folder is new, and adds the pytest.ini the tests run under. The tests need pytest: pip install pytest, then cd landing-text and python -m pytest.

Show the command
python -c "
import hashlib, pathlib, urllib.request
base = 'https://raw.githubusercontent.com/Mike-Byrne-AI/specdec-sidechannels/v1.0.0/'
out = pathlib.Path('landing-text')
files = [
    ('LICENSE', 'LICENSE', '8497cbc80c24245fc8f9be355fca4a8d15b0235b7de4870b30d573189092aeef'),
    ('tests/test_landing_text.py', 'sideband/tests/test_landing_text.py', '6db101c10b227c2e3114c86a79738067a2d72f43ac68327454224b135c487881'),
    ('tools/landing_text.py', 'sideband/tools/landing_text.py', '7bddcc8774e624d53618a029d67bd8e033ed18973110552d05f8b849e158c7aa'),
]
if out.exists():
    raise SystemExit(str(out) + ' already exists: nothing written')
got = []
for dest, src, sha in files:
    data = urllib.request.urlopen(base + src).read()
    if hashlib.sha256(data).hexdigest() == sha:
        got.append((out / dest, data))
    else:
        raise SystemExit(src + ': its sha256 is not the one on the page: nothing written')
for path, data in got:
    path.parent.mkdir(parents=True, exist_ok=True)
    path.write_bytes(data)
(out / 'pytest.ini').write_bytes(chr(10).join(['[pytest]', 'testpaths = tests', '']).encode())
print('checked and wrote', len(got), 'files to', out.resolve())
"

In the record

The record keeps the rule that numbers in prose come only from files read in that session (its CLAUDE.md working practices and docs/METHOD.md), and docs/CONVENTIONS.md records the generated-landing practice this tool served in the later lanes: every number in a generated sentence must trace to a registered payload field, and the text refuses banned phrases before it is written.

Use

  • Copy tools/landing_text.py; its Pair expects the record's payload schema (conclusion.labels, conclusion.deltas, provenance), so adapt the field names to your payloads.
  • Run python -m pytest from this folder. One test needs the record's committed R2 payloads and skips without them.

Good to know

  • This tool and the findings lint were catalogued as one registration kit together with an evidence-pin tool; the pin tool is bound to the record's own digest file and is not copied here.
The 3 files, each linked to the same file at v1.0.0
Module pathPublic file at v1.0.0sha256Bytes
LICENSELICENSE8497cbc80c241070
tests/test_landing_text.pysideband/tests/test_landing_text.py6db101c10b2221823
tools/landing_text.pysideband/tools/landing_text.py7bddcc8774e629668

critic lane

Critic lane: an adversarial reviewer agent that recomputes from state files

sideband · agent

3 files, 10201 bytes; MIT licence.

Contents: agent sideband-critic (modes: recompute, diff); skill /critic.

Tests of this copy: no tests: a prompt, not code

How to use this tool

Instructions that make Claude Code a second, sceptical reviewer for your experiments. Before you commit a result, it checks the change against a nine item checklist, marks each objection as blocking or advisory, and can recompute the claimed numbers from your saved run files instead of trusting the text.

  1. Copy the command under Get this folder and run it; the folder holds the two instruction files and the licence.
  2. Put sideband-critic.md in your project's .claude/agents/ folder and SKILL.md at .claude/skills/critic/SKILL.md.
  3. Rewrite the nine checklist items for your own work: as copied they refer to this record's decisions and findings.
  4. In Claude Code, type /critic, or ask it to run the critic on a change, a finding or a result file. Ask for recompute mode when you want the numbers checked.
  5. The agent file names a model and the tools it may use; set both for your own setup.

A Claude Code review-agent prompt and a /critic skill that check an experiment change or a finding against a nine-item checklist, tag each objection BLOCKING or ADVISORY, and in recompute mode re-derive the asserted numbers from committed state files; for researchers who run experiments with a coding agent.

Get this folder

One command makes a new folder named critic-lane inside the folder your terminal is in. It changes nothing else, joins no project and installs nothing. Paste it into bash, zsh or PowerShell with Python on the path; where python is not found, change the first word to python3. It downloads the 3 files listed below from the repository at tag v1.0.0, checks each against its full sha256 and writes nothing unless every one matches and the folder is new. The use notes say where each file goes from there.

Show the command
python -c "
import hashlib, pathlib, urllib.request
base = 'https://raw.githubusercontent.com/Mike-Byrne-AI/specdec-sidechannels/v1.0.0/'
out = pathlib.Path('critic-lane')
files = [
    ('LICENSE', 'LICENSE', '8497cbc80c24245fc8f9be355fca4a8d15b0235b7de4870b30d573189092aeef'),
    ('sideband-critic.md', 'sideband/.claude/agents/sideband-critic.md', '54b640ee2d3a46dd9ff12d4bac85b2caa7a9006e4d9cdc8e4c761293bdbddc34'),
    ('skills/critic/SKILL.md', 'sideband/.claude/skills/critic/SKILL.md', '182effe690b94f6b4228812bdbb602d0770e8f4f9986f48350c25ee7f5070993'),
]
if out.exists():
    raise SystemExit(str(out) + ' already exists: nothing written')
got = []
for dest, src, sha in files:
    data = urllib.request.urlopen(base + src).read()
    if hashlib.sha256(data).hexdigest() == sha:
        got.append((out / dest, data))
    else:
        raise SystemExit(src + ': its sha256 is not the one on the page: nothing written')
for path, data in got:
    path.parent.mkdir(parents=True, exist_ok=True)
    path.write_bytes(data)
print('checked and wrote', len(got), 'files to', out.resolve())
"

In the record

The record's workflow runs this agent before any finding is committed and before experiment changes merge; its highest-value mode recomputes the asserted labels and numbers from the committed state files rather than reading the diff. METHOD.md's reuse section names it second of its three-file kit and says to rewrite the nine checklist items for your domain while keeping the shape: named leak channels, the dominated baseline, regime abuse, survivorship, pairing, the banned claim class, kill-criteria dodging, the BLOCKING and ADVISORY tags, and the closing 'What would change my mind:'.

Use

  • Save sideband-critic.md as .claude/agents/sideband-critic.md in your project and skills/critic/SKILL.md as .claude/skills/critic/SKILL.md; then ask Claude Code to run the critic on a diff, a finding or a payload.
  • Rewrite the nine checklist items for your domain before relying on it: as copied they cite this record's decisions, findings and a kill criterion (D-2, D-5, D-6, D-7, D-9, F-3, F-5, F-6, K-5).
  • The agent file pins model: opus and a tool allowlist (the read tools plus python and git commands; the file itself says nothing mechanical keeps its use read-only); both are settings to restate for your host.

Good to know

  • The catalogue scored this piece 3: the two files drop into any Claude Code repository's .claude/ unchanged, but the agent's 'Running inside the harness' section names the record's own harness (its plan gate, scratchpad convention and a global editable install), and the dispatch paragraphs that call it live inside harness-managed command files that are not copied.
  • METHOD.md says the critic's round reports are not committed; a recovered selection of them lives in the record's docs/audits/critic-recovered/, whose README says where the reports and the dispatch briefs lived. A workflow note for a reader outside the record (briefs, reports, verbatim saves) is still to be written before this can be a card.
The 3 files, each linked to the same file at v1.0.0
Module pathPublic file at v1.0.0sha256Bytes
LICENSELICENSE8497cbc80c241070
sideband-critic.mdsideband/.claude/agents/sideband-critic.md54b640ee2d3a5675
skills/critic/SKILL.mdsideband/.claude/skills/critic/SKILL.md182effe690b93456

lockstep integer coder

Lockstep coder: a pure-integer arithmetic coder that agrees bit for bit

lockstep · folder

25 files, 98804 bytes; MIT licence.

Contents: RangeEncoder; RangeDecoder; Cdf; quantize_weights; quantize_probs; deterministic_bytes; deterministic_ints; encode_bytes; decode_bytes; encode_stream; decode_stream; roundtrip_ok; make_model; load_vectors; run_vectors; fingerprint.

Tests of this copy: 33 passed

How to use this tool

An arithmetic coder written with whole numbers only, no floating point, so an encoder on one machine and a decoder on another produce and read exactly the same bytes. It comes with frequency tables, two byte models, a seeded byte generator and six conformance vectors that check a copy or a port agrees.

  1. Copy the command under Get this folder and run it, then run python -m pytest in the new folder. It needs nothing beyond Python and pytest; conftest.py puts src on the path.
  2. Keep the package folder under src whole: its __init__.py imports the archive layer (container.py and manifest.py) as well, so using the coder alone means trimming that file.
  3. To compress bytes, call encode_bytes(data, make_model('adaptive-order1')), and keep the data's length: decode_bytes(payload, length, make_model('adaptive-order1')) needs it, with a fresh model on each side.
  4. To code your own symbols, build a table with Cdf.from_freqs([3, 1, 4]), pass each symbol's Cdf.interval to RangeEncoder.encode, then call RangeEncoder.finish. To decode, give the bytes to RangeDecoder and, for each symbol, call RangeDecoder.decode_target with the table's total, look the result up with Cdf.find, and call RangeDecoder.update with that symbol's interval.
  5. Call run_vectors() from lockstep.conformance to check this copy, or your own port, against the six pinned vectors; every result should come back ok.

A pure-integer arithmetic coder in stdlib Python with CDF quantizers, a seeded deterministic byte generator, an environment fingerprint and conformance vectors pinned by hash; for people who need an encoder and decoder to agree bit for bit across machines.

Get this folder

One command makes a new folder named lockstep-integer-coder inside the folder your terminal is in. It changes nothing else, joins no project and installs nothing. Paste it into bash, zsh or PowerShell with Python on the path; where python is not found, change the first word to python3. It downloads the 25 files listed below from the repository at tag v1.0.0, checks each against its full sha256, writes nothing unless every one matches and the folder is new, and adds the pytest.ini the tests run under. The tests need pytest: pip install pytest, then cd lockstep-integer-coder and python -m pytest.

Show the command
python -c "
import hashlib, pathlib, urllib.request
base = 'https://raw.githubusercontent.com/Mike-Byrne-AI/specdec-sidechannels/v1.0.0/'
out = pathlib.Path('lockstep-integer-coder')
files = [
    ('LICENSE', 'lockstep/LICENSE', '8497cbc80c24245fc8f9be355fca4a8d15b0235b7de4870b30d573189092aeef'),
    ('conformance/vectors_v0.json', 'lockstep/conformance/vectors_v0.json', 'c452e20b01344a9d4b1e0c0f42b17da97572de7ac52746a3c9ef6e91934b4861'),
    ('conftest.py', 'lockstep/conftest.py', '8931e98173a1dca5e11132f8db85317156bd78182ea173ce31845d3f36f81ef7'),
    ('specs/conformance-v0.md', 'lockstep/specs/conformance-v0.md', '8961c9816e7244403eb607fc7a0fee3634a3e455a8f75bade846079de3ac220c'),
    ('src/lockstep/__init__.py', 'lockstep/src/lockstep/__init__.py', 'fb19da5a4cf0dfcb9a2d510fcc080b359c56faadd59406c6cc94fa49209b13ed'),
    ('src/lockstep/cdf.py', 'lockstep/src/lockstep/cdf.py', 'dc3c9ba84bf328f4bae09f676b2c622f47d0d0f65c95a6494176586ab1cfba8f'),
    ('src/lockstep/codec.py', 'lockstep/src/lockstep/codec.py', '0f02b2215b5956d7cb8d8980e4deb94136dc9d08b88c99af362305a25778714f'),
    ('src/lockstep/conformance.py', 'lockstep/src/lockstep/conformance.py', '83639ee4664c51370724ea1c5f7ba495a97b644c8cf93d1df04c9c0c7a81c8c9'),
    ('src/lockstep/container.py', 'lockstep/src/lockstep/container.py', '81b0dc83f94c22d2f5af9d8931a28e7b8612869bcd00087b73e1282a08845703'),
    ('src/lockstep/data/__init__.py', 'lockstep/src/lockstep/data/__init__.py', '273c891b3cf71c48d52738a5455af07eae7de9c8708770e8a3c8c276d66c2930'),
    ('src/lockstep/data/vectors_v0.json', 'lockstep/src/lockstep/data/vectors_v0.json', 'c452e20b01344a9d4b1e0c0f42b17da97572de7ac52746a3c9ef6e91934b4861'),
    ('src/lockstep/drbg.py', 'lockstep/src/lockstep/drbg.py', 'c5367ba5b0306d03e30a51b8496c06eb9c674d9779d16b6580d6eb20116c6679'),
    ('src/lockstep/envprobe.py', 'lockstep/src/lockstep/envprobe.py', 'fa833d0f5be48b80a8da4272e9c11bb5cc89599e8aba3906127ced314f5c8f18'),
    ('src/lockstep/errors.py', 'lockstep/src/lockstep/errors.py', '755d94d1169565ec0aa5b7684147f2ff8b7c357db5b5cfe73f443bfdf717fc26'),
    ('src/lockstep/manifest.py', 'lockstep/src/lockstep/manifest.py', 'a00c7463d68d42e0809712a1cbbed52b33171260dea1a8ec84b7d0ae99581d3b'),
    ('src/lockstep/models.py', 'lockstep/src/lockstep/models.py', 'b40a55c06d8126ddf2f0bf697931c06d8ec4731cbf2215ee0dc1a03a47f42654'),
    ('src/lockstep/rac.py', 'lockstep/src/lockstep/rac.py', '9cc23cd7f817b6de88c3c17c52ed8f64f18aaba3cdaf9d4fd6e9132bab89a4aa'),
    ('tests_project/__init__.py', 'lockstep/tests_project/__init__.py', '8f5aebc872a6d88b5f9d29acd6df67a23d5e63cec146c42d642d18c55d7f4d78'),
    ('tests_project/test_api_validation.py', 'lockstep/tests_project/test_api_validation.py', 'ecbf0d2dec309ca06c21e2c742c8d7211c644adbea05050e11e4176754b12716'),
    ('tests_project/test_cdf.py', 'lockstep/tests_project/test_cdf.py', 'a7699d30d309d591e8223ee0a28c6ab450e36366677d4686f0673394a8ef5826'),
    ('tests_project/test_codec.py', 'lockstep/tests_project/test_codec.py', '6286642b6287b0128639db7e9ce5f011154519ef1d5faef22d8c0d51508a41e1'),
    ('tests_project/test_conformance.py', 'lockstep/tests_project/test_conformance.py', '6e0efa70663bed3e24dfb76a7169063c7ddbc9103c7e4b01ff213d516a3dda0d'),
    ('tests_project/test_packaged_data.py', 'lockstep/tests_project/test_packaged_data.py', '9719313d3436f8ead7719ff7a1e19a4cabc1f03cc1447967b6b20dddd7fe5cdd'),
    ('tests_project/test_rac.py', 'lockstep/tests_project/test_rac.py', '589358c8481a5eb244b9abcd50b3977eb6f3afc24864b69c26a4e7cd8a0631b1'),
    ('tools/lockstep_gen_vectors.py', 'lockstep/tools/lockstep_gen_vectors.py', '4840ba922d8fbc58113dbaed2a04fc968e2fd05d7751e9e5173c6dfb3ff786b9'),
]
if out.exists():
    raise SystemExit(str(out) + ' already exists: nothing written')
got = []
for dest, src, sha in files:
    data = urllib.request.urlopen(base + src).read()
    if hashlib.sha256(data).hexdigest() == sha:
        got.append((out / dest, data))
    else:
        raise SystemExit(src + ': its sha256 is not the one on the page: nothing written')
for path, data in got:
    path.parent.mkdir(parents=True, exist_ok=True)
    path.write_bytes(data)
(out / 'pytest.ini').write_bytes(chr(10).join(['[pytest]', 'testpaths = tests_project', 'pythonpath = src', '']).encode())
print('checked and wrote', len(got), 'files to', out.resolve())
"

In the record

lockstep is the third folder of the release: deterministic arithmetic coding and model-pinned archives. The coder is the part that has to agree bit for bit between an encoder and a decoder on different machines, and the conformance vectors, pinned by hash in conformance.py and shipped twice (a data copy inside the package and one beside the spec), are how that agreement is checked.

Use

  • Copy src/lockstep/ whole: the package __init__ re-exports container.py and manifest.py (the archive layer), so they come along even if you only want the coder; trimming the __init__ is the edit to make if you do not.
  • Run python -m pytest from this folder; conftest.py puts src/ on the path as the record's own layout does.

Good to know

  • rac.py describes itself as a Witten-Neal-Cleary arithmetic coder while its classes are named RangeEncoder and RangeDecoder.
  • envprobe.py's docstring points at a command-line entry point that is not copied here.
  • The two vectors_v0.json copies must stay byte-identical; the manifest records both digests.
The 25 files, each linked to the same file at v1.0.0
Module pathPublic file at v1.0.0sha256Bytes
LICENSElockstep/LICENSE8497cbc80c241070
conformance/vectors_v0.jsonlockstep/conformance/vectors_v0.jsonc452e20b01343208
conftest.pylockstep/conftest.py8931e98173a189
specs/conformance-v0.mdlockstep/specs/conformance-v0.md8961c9816e722937
src/lockstep/__init__.pylockstep/src/lockstep/__init__.pyfb19da5a4cf01828
src/lockstep/cdf.pylockstep/src/lockstep/cdf.pydc3c9ba84bf36384
src/lockstep/codec.pylockstep/src/lockstep/codec.py0f02b2215b595455
src/lockstep/conformance.pylockstep/src/lockstep/conformance.py83639ee4664c3772
src/lockstep/container.pylockstep/src/lockstep/container.py81b0dc83f94c25814
src/lockstep/data/__init__.pylockstep/src/lockstep/data/__init__.py273c891b3cf747
src/lockstep/data/vectors_v0.jsonlockstep/src/lockstep/data/vectors_v0.jsonc452e20b01343208
src/lockstep/drbg.pylockstep/src/lockstep/drbg.pyc5367ba5b0301624
src/lockstep/envprobe.pylockstep/src/lockstep/envprobe.pyfa833d0f5be42449
src/lockstep/errors.pylockstep/src/lockstep/errors.py755d94d116951287
src/lockstep/manifest.pylockstep/src/lockstep/manifest.pya00c7463d68d5862
src/lockstep/models.pylockstep/src/lockstep/models.pyb40a55c06d813693
src/lockstep/rac.pylockstep/src/lockstep/rac.py9cc23cd7f8179565
tests_project/__init__.pylockstep/tests_project/__init__.py8f5aebc872a6164
tests_project/test_api_validation.pylockstep/tests_project/test_api_validation.pyecbf0d2dec304835
tests_project/test_cdf.pylockstep/tests_project/test_cdf.pya7699d30d3092837
tests_project/test_codec.pylockstep/tests_project/test_codec.py6286642b62871983
tests_project/test_conformance.pylockstep/tests_project/test_conformance.py6e0efa70663b2201
tests_project/test_packaged_data.pylockstep/tests_project/test_packaged_data.py9719313d34361156
tests_project/test_rac.pylockstep/tests_project/test_rac.py589358c8481a3550
tools/lockstep_gen_vectors.pylockstep/tools/lockstep_gen_vectors.py4840ba922d8f3786

Glossary

The record's conclusion labels and terms in its own definitions, searchable: the 199 entries of the A to Z section of sideband/docs/GLOSSARY.md, each with its source line. The box above the list gives this page's plain words for the terms that come up most.

Plain words for the terms that come up most

Draft model, target model
The small model that proposes tokens and the large model whose output is the answer. In sideband the pair is a small Llama draft (about one billion parameters) with a larger Llama target (about eight billion); amber's pair is a Llama draft of about one billion with a target of about three billion; the regime line of each finding names them exactly.
Verify call
One forward pass of the target over a block of drafted tokens. Tokens per verify call is the record's unit for conversion: how many tokens each check emits.
The stream, telemetry, sideband
What a memory or compute device could observe about the target's forward pass: which weight slots or which experts were used. At S3 no value is seen; S1 needs the activation values to rank its units. The sideband is the narrow channel that would carry it to the drafter.
Harvest, primary and replication
A run of the target alone over the committed prompts that logs the stream and the free summaries at every block start. Each source has a primary harvest and a replication harvest on fresh seeds; a label must hold on both.
Probe, reader
A small classifier trained on the features at a block start to predict a label some positions ahead. The record's registered reader is a small multilayer network; three stronger readers were tried afterwards.
Anchor, offset
The anchor is the last verified position, whose forward produced the stream and the free summaries. The offset is how many tokens ahead the probe is asked to predict; the pinned cells use two.
Proxy baseline, full free baseline
The first reads gave the probe without the stream only the target's bucket summaries. The deciding read gave it everything free: the next token summaries, the top candidate ids and the anchor's hidden state.
Equal dims, the noise twin
To price the cost of simply adding inputs, the probe without the stream gets a block of the same size whose alignment to the text is destroyed. A margin only counts above that twin.
Gate D, materiality floor
D is the accuracy difference between the probe with the stream and the equal dims twin. The floor is the smallest D that was agreed to matter, fixed before the runs.
The bar
For conversion, the gain in tokens per verify call that was agreed to count as a win, fixed before the runs, with the rule that the interval must also clear zero.
Prompt cluster bootstrap
The interval method: resample whole prompts with replacement, recompute the statistic, take the middle 95%. Positions inside one prompt are correlated, so prompts are the unit.
Registration, pinned
In sideband, cells, readouts, bars and the conclusion labels were written into the plan, dated, before the governed result existed, with the exceptions the record names; lockstep's findings log was not registered in advance. A pinned cell is the one cell a decision was tied to before the run.
Regime
The exact conditions a number was measured under: models, quantization, block length, prompts, seeds, label space. Every finding states its regime on its own line.
Source line
Every sideband finding lists the committed files its numbers come from, and sideband's lint refuses a finding without one.
Erratum, dated note, amendment
Corrections are appended beside the text they correct, with a date, and the original is never rewritten. The timeline on this page collects them.
Mixture of experts, router
A model whose layers hold many small expert networks and send each token to a few of them. The router's choices are addresses, and at S3 they are the whole stream.
NF4
A weight format of four bits per weight, used to fit the target model on a single consumer graphics card. It is the primary regime of sideband's runs on the real models.
h_a, reconstruction
The target's internal state at the anchor, and the attempt to rebuild it from the stream with a linear map. The erratum showed the state follows from the free next token distribution.
Seed noise floor
How far the same decoder's tokens per check moves when only the random seed changes. It says what a conversion bar is worth at a given prompt count.
Fresh history export, withheld
The public repository holds the records' files as they stood, without the private commit history. What a folder leaves out is listed in its WITHHELD.md and marked in place.
Byte exact, the fp32 contract (amber)
An archive decodes byte for byte when the draft alone reproduces the stored text to the last byte. amber showed that holds under single precision arithmetic on one box and fails for decoders whose precision differs from the encoder's.
Rank corrections, bits per token (amber)
amber stores, for each token, where the true token sat in the draft's ranked guesses. The archive's size per token is the compression figure, measured against general purpose compressors.
Range coder, conformance vectors (lockstep)
A pure integer entropy coder, with fixed test inputs and their exact byte outputs, so another implementation can show it matches on what those vectors cover.
Pack rate, break-even (lockstep)
How fast an archive can be produced, against the rate at which the cost model says archiving would pay for itself. The measured rate fell far short.

The plain words explanations, the section intros and the reading guide are this page's own wording, not the record's; they add no claim and carry no chip.

A0 / A1 (Stage-A ladder)
The two user-launched rungs of the OQ-16 actuator stage (docs/BLUEPRINT.md STAGE-A FREEZE, 2026-09-05).
sideband/docs/GLOSSARY.md:215 · 6687967ae5da
ABSORBED
The strong-null label of the full-free re-tests (F-24, F-25, F-28), one of the mechanical labels per (source, label space) read from both harvests of a source: the gate D = acc(fplus_b) - acc(fplus_n) (F-25 and F-28: acc(fplus_r), the corrected twin) at the off2 gate cell (the registered reader mlp256 in F-24, F-25 and F-28's lens stage; each stronger reader in F-28's reader stages) fires on neither harvest under the Phase-3 triple rule (CI95 lower > 0, all three seeds positive, point >= +0.005) AND every D row's CI95 upper is below +0.005 on both harvests (F-24's registered rule; its siblings READER-INVALID / SURVIVES-FULL-FREE / SPLIT / NOT-SHOWN / INCOMPLETE).
sideband/docs/GLOSSARY.md:217 · 6687967ae5da
Acceptance probability, per-position (1 - TV(p, q))
Under the Leviathan rejection scheme the probability that a drafted token is accepted equals one minus the total-variation distance between the target's p and the (possibly steered) draft's q at that position.
sideband/docs/GLOSSARY.md:219 · 6687967ae5da
Actuator
The mechanism that turns a signal into acceptance: bucket-bias steering (D-8/D-10; dead as built on real models at every tested k, F-12/F-13/F-14; the fine-k oracle cell is unrun), adaptive block scheduling on probes (K-3 Attempt #1, F-19), the outcome-head scheduler (OQ-15, F-21), and SD^2-style hidden-state injection (OQ-16 Stage A, frozen, unrun).
sideband/docs/GLOSSARY.md:221 · 6687967ae5da
ACTUATOR-CEILING NULL
The Stage-A A0 outcome if oracle_h (the TRUE h_a injected) fails to beat vanilla by the E6 bar: "not even the TRUE anchor state converts through this frozen-drafter rank-512 actuator at this training budget on this box" (docs/BLUEPRINT.md STAGE-A FREEZE).
sideband/docs/GLOSSARY.md:223 · 6687967ae5da
ADDRESS-BEYOND-TOKEN
F-26's prefetch reading on the OLMoE S3 harvests: HISTORY_TOK minus FREE_TOK at horizon j = 1, in hit rate at residency budget B = 16 of the 64 experts per gate; SIGNAL iff the paired CI95 lower > 0 and the point >= +0.05 (the pinned floor), NEGATIVE iff the CI95 upper < 0, else NOT-SHOWN; "the sideband sentence follows ADDRESS-BEYOND-TOKEN alone".
sideband/docs/GLOSSARY.md:225 · 6687967ae5da
Alignment-destroyed
A feature block taken from a different block of the same prompt (slots of anchor (a+17) mod row-length, F-16), so per-prompt statistics survive but the block-aligned signal is gone.
sideband/docs/GLOSSARY.md:227 · 6687967ae5da
Analytic acceptance (D-9 headline)
E[1 - TV(p_j, q_j)] recorded at every DRAFTED position: unconditional, variance-reduced, exogenous to accept/reject outcomes.
sideband/docs/GLOSSARY.md:229 · 6687967ae5da
Analytic vs empirical vs probe units
Three layers the findings never convert between: probe top-1 accuracy margins (K-2), per-position analytic acceptance and tokens per call (K-3), and wall-clock (K-5, unmeasured).
sideband/docs/GLOSSARY.md:231 · 6687967ae5da
Anchor-state identity (F-23)
The 2026-09-05 dated erratum's fact: h_a, the verifier's final-norm output at the anchor, is the lm_head input, and Llama-3.1-8B-Instruct's lm_head is an untied, bias-free linear map of full column rank (128256 x 4096), so the anchor's logits determine h_a exactly and the full next-token distribution p determines it up to floating-point error (results/real/oq16_ha_identity.json: full-logit inverse relative error 2.5e-8 median; top-4097 log-prob recovery per-row R^2 +1.0000; top-32 R^2 +0.014 on 48 pinned rows).
sideband/docs/GLOSSARY.md:233 · 6687967ae5da
Anchor
The block-start position whose forward produced the telemetry and the free summaries.
sideband/docs/GLOSSARY.md:235 · 6687967ae5da
Argmax mechanism (argmax+tau)
D-8's original default: boost only the single argmax-predicted concept bucket, gated on confidence tau.
sideband/docs/GLOSSARY.md:237 · 6687967ae5da
ARTIFACT-DOMINANT / ARTIFACT-NOT-EXCLUDED
DI-1's pinned labels: primary stale point >= 0.5 x the live margin -> ARTIFACT-DOMINANT; the replicate below that -> ARTIFACT-NOT-EXCLUDED; disagreement carries ARTIFACT-DOMINANT.
sideband/docs/GLOSSARY.md:239 · 6687967ae5da
Attempt #1 (K-3)
The registered Phase-4 arc that first tested conversion: adaptive block scheduling on frozen S1 probes, five arms, thresholds frozen on 30 val prompts, verdict on data/prompts_k3_verdict.jsonl (210 prompts).
sideband/docs/GLOSSARY.md:241 · 6687967ae5da
B_ref
The reference drafting budget: vanilla gamma-4's drafted-per-emitted measured on the VAL prompts, frozen at first read before any verdict run; each adaptive arm tunes subject to val dpe <= B_ref.
sideband/docs/GLOSSARY.md:245 · 6687967ae5da
BAR-RESOLVED@210 / RULE-LEAKS@30 (F-27 readings)
The seed-noise floor's two readings beside its label, each mechanical (docs/BLUEPRINT.md, "SEED-NOISE FLOOR READOUTS REGISTRATION"): BAR-RESOLVED@210 iff the n = 210 cross-seed null CI95 half-width is below 0.062 (the +0.062 WIN bar exceeds one CI half-width at the verdict size), else BAR-UNRESOLVED@210; RULE-LEAKS@30 iff g_one at n = 30, the share of 1,000 prompt subsets in which a single comparator leg of the F-19 WIN rule passes by seed noise alone, exceeds 0.01, else RULE-TIGHT@30 ("1% is a reporting threshold with NO consequence: the rate itself is the datum and is quoted with its SE whatever the word").
sideband/docs/GLOSSARY.md:247 · 6687967ae5da
Beta
Steering strength, the magnitude of the additive log-space bias.
sideband/docs/GLOSSARY.md:249 · 6687967ae5da
Binding gate
See Sha header / binding gate.
sideband/docs/GLOSSARY.md:251 · 6687967ae5da
Block
One draft-then-verify round: gamma drafted tokens, one target verify forward, emitted = accepted + 1 (the bonus or residual token).
sideband/docs/GLOSSARY.md:253 · 6687967ae5da
Block-aligned information vs prompt-level statistics
The distinction the stale-target control draws: information tied to THIS block's anchor (what steering needs) versus information about the prompt overall (which a shifted twin also carries).
sideband/docs/GLOSSARY.md:255 · 6687967ae5da
Block-start alignment
The load-bearing harvest convention (sideband.harvest.label_index): the label at offset j is the bucket of token x_{a+j-1}; offset 1 is the token sampled from p_a itself; K-2 reads offsets >= 2.
sideband/docs/GLOSSARY.md:257 · 6687967ae5da
Bonus token / residual token
The one token the verifier always emits after a block: sampled from the residual distribution on a rejection, or from the target's distribution at position gamma+1 when every draft is accepted.
sideband/docs/GLOSSARY.md:259 · 6687967ae5da
Break-even hint budget
F-9 (synthetic, OQ-7): the total bits per token an archive hint channel may spend before it costs more than the corrections it saves - 0.162 bits/token in that regime; a dense per-position fresh-concept hint loses 20:1 (net -3.18 bits/token); free-causal steering saves 0.100 bits/token at zero cost.
sideband/docs/GLOSSARY.md:261 · 6687967ae5da
Bucket / bucketization
A partition of the vocabulary.
sideband/docs/GLOSSARY.md:263 · 6687967ae5da
Byproduct logs
E8: per K-3-arc run, a gitignored directory <out>.byproduct/ with one npz per prompt holding per-block arrays (anchor position, slots, h_a, bucket masses at several k, entropy, argmax, top-32 ids and probs, gamma used, stop/extend, accepted count, consult signals) and per-drafted-position arrays (proposed token, accepted flag, analytic 1 - TV).
sideband/docs/GLOSSARY.md:265 · 6687967ae5da
C1..C4 (OQ-15 Stage-C arms)
C1 vanilla (runner arm 1); C2 draft-confidence (arm 2, the SpecDec++-class incumbent retuned on fresh val); C3 head_free = the outcome-head policy on FREE features, 657 dims (arm 6, the matched-actuator free baseline); C4 head_tel = the same policy on FREE+TEL, 2705 dims (arm 7, the bet).
sideband/docs/GLOSSARY.md:269 · 6687967ae5da
Ceiling vs existence vs headroom vs actuator
Four layers the findings keep apart.
sideband/docs/GLOSSARY.md:271 · 6687967ae5da
Chat template, pinned date
K-1 amendment (docs/WHY_THIS_DIES.md): Instruct models are decoded with the chat template applied and a fixed date_string, recorded as chat_template_used in every run meta, because raw-completion decoding of an Instruct pair is a regime nobody deploys.
sideband/docs/GLOSSARY.md:273 · 6687967ae5da
CI-clear / CI-positive / CI-negative
A CI95 entirely above zero (clear or positive) or entirely below zero (negative).
sideband/docs/GLOSSARY.md:275 · 6687967ae5da
Committed evidence (results/*.json)
Result files are regenerated by tools, never hand-edited (CLAUDE.md).
sideband/docs/GLOSSARY.md:277 · 6687967ae5da
Composition confound
Why analytic acceptance is reported-not-gating for scheduling arms: an arm that stops aggressively drafts an easier mix of positions and posts higher per-drafted-position acceptance without producing more tokens per call.
sideband/docs/GLOSSARY.md:279 · 6687967ae5da
Compute-tier tap vs address stream
D-1's bright line.
sideband/docs/GLOSSARY.md:281 · 6687967ae5da
Concept bucket / concept space / concept-level
Non-glue buckets.
sideband/docs/GLOSSARY.md:283 · 6687967ae5da
Concept-HMM v0 (synthetic world)
The Phase-1 testbed: a hidden-Markov concept process emitting bucketed tokens, a corrupted draft (draft_alpha), concept persistence rho, a function-word rate lambda_function, and hardened telemetry parameters.
sideband/docs/GLOSSARY.md:285 · 6687967ae5da
Confidence gate / tau
D-8: steer only when the decoder's confidence exceeds tau (about 0.5); it converts harmful weak decoders into harmless ones (F-5: at the three narrowest widths of the sweep, K = 8/16/32 slots per experiments/e05_sidecar_recovery.py, recovery goes from -152%/-88%/-22% ungated to -1%/+2%/+8% gated).
sideband/docs/GLOSSARY.md:287 · 6687967ae5da
Constrained maximization / frontier search
The registered tuning rule for adaptive arms: maximize val tokens/call subject to val drafted-per-emitted <= B_ref over a pinned quantile grid, tie-break toward lower budget - never unconstrained tokens/call maximization (docs/ECONOMICS.md section 5 amendment 1).
sideband/docs/GLOSSARY.md:289 · 6687967ae5da
Consult / STOP / EXTEND
The K-3 and OQ-15 scheduler's per-block decisions.
sideband/docs/GLOSSARY.md:291 · 6687967ae5da
Containment rider (E10)
The empirical READ-ONLY check that instruments do not perturb the trajectory: vanilla arm, 10 val prompts, seed 100, three runs (instruments ON, OFF, OFF rerun).
sideband/docs/GLOSSARY.md:293 · 6687967ae5da
Contamination
The test-in-train prompt-overlap audit of a harvest (results/real/harvest_audit*.json): 0.8% at S1 (OPEN_QUESTIONS.md OQ-1); 0.52%/0.64% (distinct/occurrence) on the S3 primary harvest and 0.37%/0.47% on its replication (F-17).
sideband/docs/GLOSSARY.md:295 · 6687967ae5da
Coupling (about 6.2 tokens/call per unit acceptance)
docs/ECONOMICS.md section 2: a linearization from ONE dose point on the harm side of one arm (F-12 sharp beta4: -0.0237 acceptance, -0.146 tokens/call), labeled an approximation wherever used, and out-of-regime for scheduling arms.
sideband/docs/GLOSSARY.md:297 · 6687967ae5da
Critic (sideband-critic) / recompute
The .claude/agents/ review agent run against every experiment diff before merge (CLAUDE.md).
sideband/docs/GLOSSARY.md:299 · 6687967ae5da
Cross-fit h_hat (Stage-A export)
In the Stage-A freeze, the reconstructed vector used to TRAIN an injection adapter comes from five-fold cross-fitting per population (position mod 5): every prompt's h_hat comes from a ridge fit that never saw it, on 168 of 210 prompts; a full refit on all 210 yields the verdict-time weights; every fit at the arm's DR-4 VAL-chosen lambda with DR-1's support clip (docs/BLUEPRINT.md STAGE-A FREEZE).
sideband/docs/GLOSSARY.md:301 · 6687967ae5da
Cross-lane skip condition
From the FOLD-BACK COUNT AMENDMENT (2026-09-02): if OQ-15 nulls at the information ceiling, OQ-16 is SKIPPED and the lane stops, because both lanes read the same stream.
sideband/docs/GLOSSARY.md:303 · 6687967ae5da
D2/D3 (K-3 diagnostic stage)
The Attempt-#1 stage on the committed S1 harvest that produced the OQ-14(ii) error-predictor and OQ-14(iii) calibration readouts (results/real/error_predictor_s1.json: test precision@0.20 = 0.7919 (a) / 0.8356 (c) over base rates 0.6059/0.6223; test ECE 0.0082 (a) / 0.0408 (c)) and thereby picked the live probes' label space (L1/off2) and confidence signal (the raw posterior).
sideband/docs/GLOSSARY.md:307 · 6687967ae5da
data/prompts.jsonl (210)
The committed short-prompt set (one unique topic per prompt so bootstrap clusters are independent) used from Phase 0a through Phase 3; it includes the 150 probe-training prompts, which is why every actuator verdict uses a fresh committed population.
sideband/docs/GLOSSARY.md:309 · 6687967ae5da
Dated amendment / dated erratum
House rule: registered text is never rewritten; a correction is appended with a date (examples: F-16's 3266 -> 3265 dims erratum; E12(iii) "meta" -> "summary"; DR-4's "dense-valued" -> sparse TOK_B).
sideband/docs/GLOSSARY.md:311 · 6687967ae5da
Decision (D-N)
A standing decision in DECISIONS.md (D-1..D-11); see the Decisions section above.
sideband/docs/GLOSSARY.md:313 · 6687967ae5da
Defined comparison (E2)
A probe-vs-draft class comparison counts only when BOTH classes are concept classes (neither is the merged glue class); an undefined comparison fires neither STOP nor EXTEND.
sideband/docs/GLOSSARY.md:315 · 6687967ae5da
delta_noise / delta_free
The OQ-15 and OQ-16 gate deltas on pooled TEST MSE: delta_noise = MSE(NOISE) - MSE(TEL) (existence above dilution); delta_free = MSE(FREE) - MSE(TEL) (deployability above free).
sideband/docs/GLOSSARY.md:317 · 6687967ae5da
DI-1
The Stage-I stale-target control; see Stale-target control.
sideband/docs/GLOSSARY.md:319 · 6687967ae5da
Dilution
The accuracy or MSE cost of adding uninformative dimensions to a linear model at fixed data.
sideband/docs/GLOSSARY.md:321 · 6687967ae5da
Direction proxy
The OQ-16 registration's downstream readout (does the reconstructed vector's direction correlate with the true one's), pinned REPORTED-only, not gating, as one of the three design decisions put to the user before registration (docs/BLUEPRINT.md OQ-16 ARC REGISTRATION).
sideband/docs/GLOSSARY.md:323 · 6687967ae5da
Dirty flag (git_commit -dirty)
Meta provenance flag: the working tree had uncommitted files when the run started.
sideband/docs/GLOSSARY.md:325 · 6687967ae5da
Disagreement gate
Apply steering only where the draft's own top CONCEPT bucket differs from the predicted bucket, glue excluded on both sides (D-7, D-8).
sideband/docs/GLOSSARY.md:327 · 6687967ae5da
DR-1..DR-5 (Stage-R post-gate diagnostics)
Registered after the Stage-R gate was read and before running, with pinned labels; NON-CRITERION (docs/BLUEPRINT.md; F-22).
sideband/docs/GLOSSARY.md:329 · 6687967ae5da
DR labels (TELEMETRY-BEYOND-P, TOKEN-IDENTITY-NOT-EXCLUDED, DEPLOYABLE-MARGIN CONFIRMED, COMPARATOR-TOOTHLESS)
The pinned outcomes of DR-3, DR-2, DR-1 and DR-4 respectively (results/real/oq16_stage_r_dr4.json diagnostic_labels).
sideband/docs/GLOSSARY.md:331 · 6687967ae5da
draft_alpha / rho
Synthetic-world knobs: draft_alpha = draft quality (0.5 weak .. 0.9 strong; F-2 cross-bucket fractions 89.3% / 83.1% / 76.1%); rho = concept persistence (run length 2.5 at rho=0.6, 33 at rho=0.97; F-4).
sideband/docs/GLOSSARY.md:333 · 6687967ae5da
Drafted-per-emitted (dpe)
E6: drafted tokens divided by emitted tokens - the drafting BUDGET an arm spends per token it produces.
sideband/docs/GLOSSARY.md:335 · 6687967ae5da
E1..E12 (Stage-E engineering clarifications)
Twelve pins in docs/BLUEPRINT.md (dated 2026-09-02, before any stage-E code) closing the degrees of freedom of the K-3 live instrument: (E1) live probes = the four committed Phase-3 cells L1_concept_k16 | off{2,3} | {a,c} | linear at train seed 0, binding-gated export; (E2) agreement/disagreement semantics in concept space (see Defined comparison); (E3) arm-2 signal mapping signal_j = q_j(x_j); (E4) extend mechanics (one decision per block after the base drafting); (E5) threshold materialization from quantile grids over consulting blocks; (E6) gate arithmetic (tokens/call = emitted per verify call, prefill excluded; the joint-bootstrap ratio delta); (E7) feature fidelity to the harvest dtypes; (E8) byproduct format; (E9) live-tap fidelity rider + tolerance mapping; (E10) containment rider; (E11) instrument change scope (defaults OFF, byte-identical default path); (E12) dated amendment: block-1 consult exclusion, temperature binding, the E8 erratum (shas in the summary, not the meta), E9 errata.
sideband/docs/GLOSSARY.md:339 · 6687967ae5da
Empirical acceptance-by-offset (survival-conditioned)
The fraction of drafts accepted at offset j among blocks that SURVIVED to offset j (F-10 at k=512: 0.815/0.818/0.809/0.807 for offsets 1-4, n = 4321/3015/2310/1641).
sideband/docs/GLOSSARY.md:341 · 6687967ae5da
Equal-dims noise anchor / noise twin / NOISE / (atn)
The control that prices dilution: the same base features plus a telemetry block whose ALIGNMENT is destroyed, so dims match the tested arm exactly (F-16: (atn) = (a)+(t)+shifted slots, 3265 dims = (abt)).
sideband/docs/GLOSSARY.md:343 · 6687967ae5da
EQUAL-DIMS-ONLY
The licence qualifier of a SURVIVES statement in the full-free re-tests, pinned in the OQ-18 registration's LICENSED SENTENCES (docs/BLUEPRINT.md): a SURVIVES statement quotes Mt AND Mn_r beside D and names the reader; the licence is FULL only if BOTH Mt's and Mn_r's CI95 uppers are >= 0 at the gate cell on both harvests, since by the paired identity D = Mt - Mn_r a gate fire on a harvest where the corrected twin is CI-clear harmful is bought by the twin's cost; a fire with Mn_r's upper < 0 is written "at equal dims (D), inflated by up to |Mn_r| by the twin's own cost" and a fire with Mt's upper < 0 is written "at equal dims (D); at this budget adding it did not raise accuracy (Mt)" (the payload's survives_license and license_reason say which).
sideband/docs/GLOSSARY.md:345 · 6687967ae5da
Existence
The first of the three layers; see The three layers and K-2.
sideband/docs/GLOSSARY.md:347 · 6687967ae5da
Extend-signal breakdown
F-21 finding duty (i): the offset-5 heads that gate every EXTEND land WORSE than the trivial train-mean baseline on TEST (free 0.05469985821808974 / tel 0.05974761296262184 vs trivial 0.052217), so the registered actuator was mostly steering on noise wherever it extended; the measured null prices the actuator WITH that channel, and a stop-only variant was never registered.
sideband/docs/GLOSSARY.md:349 · 6687967ae5da
Feature sets (a), (b), (c), (h), (t), (at), (abt), (atn)
Phase-3 probe inputs (F-15, F-16).
sideband/docs/GLOSSARY.md:353 · 6687967ae5da
Finding (F-N)
An entry in FINDINGS.md; see Prefixes.
sideband/docs/GLOSSARY.md:355 · 6687967ae5da
Firing conditions (K-2 probe rule)
A triple FIRES iff (c)-(a) has CI95 lower > 0 AND all three train seeds are positive AND the point is >= +0.005 (the materiality floor) (F-15).
sideband/docs/GLOSSARY.md:357 · 6687967ae5da
FLOOR-MATERIAL
F-27's label per population (docs/BLUEPRINT.md, "SEED-NOISE FLOOR READOUTS REGISTRATION"): FLOOR-MATERIAL iff f_one at n = 30 is at or above 0.02, the pinned one-sided kill threshold, else FLOOR-WEAK; f_one is the share of 1,000 prompt subsets in which a +0.062 point win appears by a same-policy seed change alone.
sideband/docs/GLOSSARY.md:359 · 6687967ae5da
Fold-back / fold-back #1 / fold-back #2
What K-3's kill text prescribes after a null: try a mechanistically different actuator, once.
sideband/docs/GLOSSARY.md:361 · 6687967ae5da
Free information / free features / FREE
Everything a speculative decoder already has at the last verified position: the target's distribution p_a and its summaries (bucket masses, entropy, top-32 ids and probs) and, in live scheduling stages, the draft's own confidence q_sig.
sideband/docs/GLOSSARY.md:363 · 6687967ae5da
free_lastp
The D-5 baseline policy: steer the draft from the target's last verified distribution p_last via its bucket masses.
sideband/docs/GLOSSARY.md:365 · 6687967ae5da
Free actuator / FREE-ACTUATOR WIN / NULL
After F-23, the only live question the Stage-A instrument can still answer: does injecting the verifier's OWN anchor state h_a (computable from p; arm 8, code name oracle_h) into the frozen drafter convert to tokens per call on this box?
sideband/docs/GLOSSARY.md:367 · 6687967ae5da
FREETOK
DR-2's equal-dims, deployment-free comparator: FREE ++ a 2048-dim multi-hot of the anchor's top-32 next-token ids hashed mod 2048 (about 62 ids per bin, identity partly lost; 2705 dims).
sideband/docs/GLOSSARY.md:369 · 6687967ae5da
FREETOK2
See TOK_A / TOK_B / FREETOK2.
sideband/docs/GLOSSARY.md:371 · 6687967ae5da
Gamma (block length)
The number of draft tokens proposed per verify call.
sideband/docs/GLOSSARY.md:375 · 6687967ae5da
Gate-open rate
The fraction of steerable positions at which the disagreement gate fires.
sideband/docs/GLOSSARY.md:377 · 6687967ae5da
Glue bucket
Function-token buckets (", the, ., a, and, of, to, 's", markdown markers, pronouns per F-11).
sideband/docs/GLOSSARY.md:379 · 6687967ae5da
Glue tax / function-word tax / F-3 tax
The self-cancellation of ungated bucket boosting: wrong-concept positions gain, already-right positions pay because boosting a bucket steals mass from glue tokens.
sideband/docs/GLOSSARY.md:381 · 6687967ae5da
Granularity trap
F-14's summary of the bucket-bias actuator: fine k has headroom (0.61 concept-cross at k=512) but a useless gate and stale forecasts; coarse k has a working gate but true concept disagreement collapses to about 16% and the ceiling is <= +0.0069.
sideband/docs/GLOSSARY.md:383 · 6687967ae5da
Grid-edge
See Ridge / lambda / grid-edge duty.
sideband/docs/GLOSSARY.md:385 · 6687967ae5da
h_a
The anchor hidden state at the position that produced p_a: the model.norm output, 4096 dims, float16 on disk, at S1 (F-22); the S3 harvest's hidden reference feature (h) is 2048-dim (F-17), its capture point not stated in the finding.
sideband/docs/GLOSSARY.md:389 · 6687967ae5da
H-ABSORBS
R2's hidden-state (Q2) label in the OQ-9 prefetch rungs (docs/BLUEPRINT.md, the R2 registration; F-26's status lines of 2026-10-03 and 2026-10-04): H-SURVIVES iff D_eq^H and D_add^H both read SIGNAL, H-COMPLEMENTARY iff only D_add^H does, H-ABSORBS iff neither does, H-INCONSISTENT iff only D_eq^H does; a NEGATIVE reading is "not SIGNAL" and is said in those words.
sideband/docs/GLOSSARY.md:391 · 6687967ae5da
HANDOFF.md
An untracked, gitignored session-state file, not committed evidence; its most recent note records the session state; no Stage-A GPU run exists.
sideband/docs/GLOSSARY.md:393 · 6687967ae5da
Hard stop
After fold-back #2 (or the skip) no third mechanism exists or may be registered (FOLD-BACK COUNT AMENDMENT).
sideband/docs/GLOSSARY.md:395 · 6687967ae5da
Hardened telemetry
Synthetic telemetry-noise parameters (support_gain=0.5, support_noise=0.5, background_noise=0.25, top_n=16) used in e05/e06/e07 (F-5, F-7, F-9) so that sidecar decoding is non-trivial.
sideband/docs/GLOSSARY.md:397 · 6687967ae5da
Harvest
The Phase-2 protocol: run the TARGET alone (no draft) on the 210 prompts across generation seeds, logging per anchor the telemetry, the free summaries of p_a, h_a and the generated tokens, block-start aligned, with a volume bar of >= 1M (anchor, offset) pairs for a full harvest.
sideband/docs/GLOSSARY.md:399 · 6687967ae5da
Horizon
How many tokens ahead telemetry demonstrably carries information: 2-3 at S1 (F-15; offset 4 not established); at S3 offset 2 in L1, 3 in L2, 4 in L3 (F-17).
sideband/docs/GLOSSARY.md:401 · 6687967ae5da
Information-ceiling null vs conversion-mode null
OQ-15's two null branches (FOLD-BACK COUNT AMENDMENT decidability pin; OQ-15 arc registration).
sideband/docs/GLOSSARY.md:405 · 6687967ae5da
Information-free extend channel / gamma inflation
Raw tokens/call rises with gamma regardless of information (emitted = accepted + 1), so any always-extend arm clears a raw tokens/call bar; docs/ECONOMICS.md section 5 amendment 1 introduced the matched-budget frame and the pairwise guard to block it, and F-19's gamma-6 control confirmed the arithmetic (about +0.41 tokens/call, seed-1 pooled 3.428206 vs vanilla 3.020427, for +0.43 dpe).
sideband/docs/GLOSSARY.md:407 · 6687967ae5da
Invariance leg
F-22's requirement that the extended Stage-R tool with all diagnostic modes OFF reproduces the registered numbers: it held on every gate-bearing field, with the REPORTED trivial MSE differing by one ULP (0.9900697257801255 vs 0.9900697257801256) for a summation-order reason explained per the pin (results/real/oq16_stage_r_inv.json).
sideband/docs/GLOSSARY.md:409 · 6687967ae5da
Isolated RNG stream
D-6: a policy that reads telemetry draws from its own generator so reading cannot perturb the shared draft/verify trajectory.
sideband/docs/GLOSSARY.md:411 · 6687967ae5da
K-2-ISOLATED
The registered K-2 word at S3 over STAGE READERS and STAGE READERS-LONG of F-28's reader-and-lens re-test (docs/BLUEPRINT.md, "READER-AND-LENS RE-TEST AT S3 REGISTRATION"): K-2-ISOLATED iff some valid (reader, space) cell reads SURVIVES-STRONGER but the REOPENED condition fails (not reopen-grade, or at the lopsided lens, or S1 not reopen-grade, or S1 not yet run), where a cell is reopen-grade iff SURVIVES-STRONGER with the FULL licence AND Mt firing the Phase-3 triple rule on both harvests; its siblings: K-2-NOT-OVERTURNED-AT-S3 (every valid (reader, space) cell reads ABSORBED or NOT-SHOWN), K-2-UNSETTLED (some valid cell reads SPLIT and none SURVIVES-STRONGER), K-2-INCOMPLETE (no valid cell exists) and K-2-REOPENED-AT a (reader, space) cell (the reopen branch); the lopsided lens never decides a reopen.
sideband/docs/GLOSSARY.md:415 · 6687967ae5da
K-2-NOT-OVERTURNED-AT-S1
The registered S1 word of F-29's S1 reader re-test (docs/BLUEPRINT.md, "S1 READER RE-TEST REGISTRATION", 2026-10-16), issued by tools/run_reader_s1.py over the valid readers at the S1 gate cell from the frozen per-reader labels: K-2-INCOMPLETE-AT-S1 iff no valid reader; K-2-REOPENED-AT-S1 iff some valid reader is reopen-grade (SURVIVES-STRONGER with the FULL licence and Mt firing the triple rule on both harvests), per source, PROVISIONAL until a reader-matched rider passes; K-2-ISOLATED-AT-S1 iff some valid reader reads SURVIVES-STRONGER without reopen grade; K-2-UNSETTLED-AT-S1 iff some valid reader reads SPLIT and none SURVIVES-STRONGER; K-2-NOT-OVERTURNED-AT-S1 iff every valid reader reads ABSORBED or NOT-SHOWN (narrowed where ABSORBED, not where NOT-SHOWN); precedence in that order.
sideband/docs/GLOSSARY.md:417 · 6687967ae5da
K-3 arms 1-5
Attempt #1's arms (F-19): 1 vanilla gamma-4; 2 draft-confidence stop/extend (the incumbent-style comparator; selected stop q0.25 / extend q0.5); 3 the (a)-probe free-signal scheduler (q0.95 / q0.05, degenerate tuning); 4 the (c)-probe telemetry scheduler (the bet; q0.4 / q0.05, re-frozen q0.35 / q0.05 under the void); 5 fixed gamma-6 (information-free control, never in the gate minimum).
sideband/docs/GLOSSARY.md:435 · 6687967ae5da
Kill criterion (K-N)
A pre-registered death condition in docs/WHY_THIS_DIES.md; see the Kill criteria section above.
sideband/docs/GLOSSARY.md:437 · 6687967ae5da
KV cache / --no-kv-cache rider / KV-crop fast path
Real runs use per-prompt KV caches; a pre-registered cached-vs-uncached A/B must pass bidirectionally (six containments) before cached numbers are cited (K-1 rider; F-10, F-12, F-14).
sideband/docs/GLOSSARY.md:439 · 6687967ae5da
Label spaces L1 / L2 / L3
Probe target spaces (F-15): L1 = concept_k16 (16 k-means buckets, merged glue class included); L2 = concept_k512 (510 concepts + a merged glue class); L3 = concept_groups_v1 (transition-coherent groupings of the k512 buckets, fit on the train split only; a lopsided "mega-group lens").
sideband/docs/GLOSSARY.md:443 · 6687967ae5da
Lane A / Lane B
docs/ECONOMICS.md section 4's deployment lanes.
sideband/docs/GLOSSARY.md:445 · 6687967ae5da
LEAD-TIME-ABSORBED
The horizon read's stage label (docs/BLUEPRINT.md, "OQ-9 STAGE-1 HORIZON READ REGISTRATION"; F-26's status line of 2026-10-04; K-6's lead-time datum): LEAD-TIME-ABSORBED iff H-ABSORBS at every horizon j in {2, 3, 4}; LEAD-TIME-SURVIVES@j iff H-SURVIVES at some j, j the smallest such horizon; LEAD-TIME-COMPLEMENTARY@j iff no H-SURVIVES and H-COMPLEMENTARY at some j; LEAD-TIME-INCONSISTENT otherwise; the primary harvest's label is the label, the replication computed identically and reported as agrees / disagrees.
sideband/docs/GLOSSARY.md:447 · 6687967ae5da
Leviathan rejection scheme / spec_step
The standard speculative-decoding verifier: accept draft token x with probability min(1, p(x)/q(x)); on rejection sample from the residual normalize(max(p - q, 0)).
sideband/docs/GLOSSARY.md:449 · 6687967ae5da
Live-tap fidelity rider (E9) / tolerance mapping
Compares live-captured slots, mass, entropy and argmax at sampled anchors against a fresh teacher-forced recompute, with constants derived mechanically from a calibration run: SLOT_OVERLAP_MIN = 20, MASS_LINF_MAX = 0.11094206571578979, ENTROPY_ABS_MAX = 0.3908865451812744, ARGMAX_GAP = MASS_LINF_MAX (docs/BLUEPRINT.md STAGE-E TOLERANCE AMENDMENT).
sideband/docs/GLOSSARY.md:451 · 6687967ae5da
Llama pair
Llama-3.1-8B-Instruct target (NF4) + Llama-3.2-1B-Instruct draft (bf16), one shared tokenizer, gamma 4, temperature 1.0, chat template with a pinned date - the regime of every real steering and actuator finding (F-10 onward).
sideband/docs/GLOSSARY.md:453 · 6687967ae5da
M2 / M3 / M4 / M2n / M2noise
The P1 readouts (F-16, F-18).
sideband/docs/GLOSSARY.md:457 · 6687967ae5da
Majority floor
The accuracy of always predicting the most common label - the baseline a probe must lift off.
sideband/docs/GLOSSARY.md:459 · 6687967ae5da
Matched drafting budget
The K-3 gate frame (docs/ECONOMICS.md section 5 amendment 1): compare tokens/call only between arms spending the same drafting budget, because raw tokens/call is buyable information-free by drafting more.
sideband/docs/GLOSSARY.md:461 · 6687967ae5da
Materiality floor (+0.005)
The registered minimum probe-margin point estimate for a triple to count (F-15 firing conditions).
sideband/docs/GLOSSARY.md:463 · 6687967ae5da
Mechanism-matched
D-11: a ceiling or baseline must run the same mechanism as the policy it prices.
sideband/docs/GLOSSARY.md:465 · 6687967ae5da
Mega-group lens
The caveat that L3 (concept_groups_v1) is lopsided: two groups cover about 455 of 510 concepts at S1 (F-15) and about 461 of 510 at S3 (F-17, leading sizes 193 and 268), so L3's larger margins read through a coarse lens.
sideband/docs/GLOSSARY.md:467 · 6687967ae5da
n_steerable
The number of drafted positions at which a gated policy could fire, the denominator of the gate-open rate (F-14: 69,192 for the oracle arm); the free arm's count excludes each prompt's first block, the oracle's includes it (F-14 comparability footnote).
sideband/docs/GLOSSARY.md:471 · 6687967ae5da
NF4
4-bit NormalFloat quantization (bitsandbytes; --load-4bit, double-quant, bf16 compute): the primary real-model regime for the 8B target on the 12 GB RTX 3080 Ti.
sideband/docs/GLOSSARY.md:473 · 6687967ae5da
NON-CRITERION
A registered diagnostic that reports and prices but cannot change a verdict or partition branch (P1, DI-1, DR-1..DR-4, OQ-13 items).
sideband/docs/GLOSSARY.md:475 · 6687967ae5da
NOT WIN / WIN / NULL
Gate outcome labels.
sideband/docs/GLOSSARY.md:477 · 6687967ae5da
NOT-SHOWN
A conclusion label in two registered families, distinct from the status word "not shown" in the table at the top.
sideband/docs/GLOSSARY.md:479 · 6687967ae5da
Observability constraint
The premise that only addresses (widened by D-1's scope note to "almost nothing") cross the wire - no activation values, no logits.
sideband/docs/GLOSSARY.md:483 · 6687967ae5da
Offset
Distance ahead of the anchor, 1..gamma.
sideband/docs/GLOSSARY.md:485 · 6687967ae5da
OLMoE
allenai/OLMoE-1B-7B-0125-Instruct: the S3 source model, 16 MoE gates each routing top-8-of-64 experts; harvested target-only, routers and lm_head excluded from quantization (F-17).
sideband/docs/GLOSSARY.md:487 · 6687967ae5da
Open question (OQ-N)
An entry in OPEN_QUESTIONS.md; see the Open questions section above.
sideband/docs/GLOSSARY.md:489 · 6687967ae5da
Operational bar (+0.044 acceptance)
The Lane-B bar (docs/ECONOMICS.md section 5): point acceptance delta >= +0.044, about +0.27 tokens/call via the coupling, 8.8% of the 3.11 base; like-for-like on SD^2's accepted-tokens base about 37% of SD^2's gain, harder than its "25% of SD^2" motivating label.
sideband/docs/GLOSSARY.md:491 · 6687967ae5da
Oracle (stale / fresh / oracle_nextp / oracle_h / oracle_concept_now)
D-2-named non-causal CEILINGS.
sideband/docs/GLOSSARY.md:493 · 6687967ae5da
Outcome head / a_hat
OQ-15's closed-form ridge regressor predicting per-position analytic acceptance from anchor features; a_hat_j is its prediction at offset j.
sideband/docs/GLOSSARY.md:495 · 6687967ae5da
P(k=0..4) accepted-length distribution
From the committed phase0b vanilla records: 0.2509/0.1751/0.1341/0.0953/0.3445 over 17,220 blocks, E[emitted per call] 3.1071 (docs/ECONOMICS.md section 5; F-19 CONTROL ARITHMETIC) - the evidence that the information-free extend channel is real.
sideband/docs/GLOSSARY.md:499 · 6687967ae5da
P1 / P1-analog
The registered post-hoc diagnostic testing whether a probe margin is merely prefix TOPIC STATE: feature set (t) = 560 decoder-free dims of concept histograms over already-generated tokens; readouts M2/M3/M4/M2n/M2noise; pinned labels P1-SURVIVES vs P1-DEFLATES.
sideband/docs/GLOSSARY.md:501 · 6687967ae5da
P-stage
The Phase-4 PRECONDITION run of K-3 Attempt #1: arms 1-3 on the verdict population at the verdict seed, evaluated BEFORE the telemetry arm; had the free-signal arm (3) beaten vanilla the arc would have folded back early.
sideband/docs/GLOSSARY.md:503 · 6687967ae5da
p_a / p_last / last-p
The target's next-token distribution at the anchor (last verified position).
sideband/docs/GLOSSARY.md:505 · 6687967ae5da
Paired / identical seeds
D-6: arms run on the same prompts with the same per-prompt RNG streams so deltas are within-prompt differences.
sideband/docs/GLOSSARY.md:507 · 6687967ae5da
Pairwise budget guard
The verdict-time tolerance: the tested arm's verdict-population dpe must be <= each comparator's dpe + 0.01 (epsilon pinned; the information-free skew residual measured at +0.0103 tokens/call, 17% of the bar; docs/ECONOMICS.md section 5).
sideband/docs/GLOSSARY.md:509 · 6687967ae5da
Persistence
See Staleness / persistence / run length / hold rule.
sideband/docs/GLOSSARY.md:511 · 6687967ae5da
Phases 0a / 0b / 0c / 1 / 2 / 3 / 4 / 5
docs/BLUEPRINT.md's plan: 0a the real cross-bucket audit (K-1; F-10, F-11); 0b the free-information steering baseline (F-12); 0c coarse gating / OQ-12 plus the attribution rider (F-13, F-14); 1 the synthetic testbed (F-2..F-9, with F-0/F-1 the verifier-exactness floor); 2 the real telemetry harvest; 3 the existence test / K-2 (F-15..F-18); 4 steered speculative decoding end to end / K-3 (F-19..F-22 and Stage A); 5 the hardware-premise memo (docs/ECONOMICS.md skeleton).
sideband/docs/GLOSSARY.md:513 · 6687967ae5da
posterior_bias (centered posterior mechanism)
D-10's default: bias concept bucket b by beta*(P(b) - 1/C), glue 0 (sideband.steering.posterior_bias).
sideband/docs/GLOSSARY.md:515 · 6687967ae5da
PREFETCH-HEADROOM
F-26's stage label on the primary OLMoE S3 harvest (docs/BLUEPRINT.md, "OQ-9 STAGE-1 PREFETCH READOUTS REGISTRATION"): PREFETCH-HEADROOM if not CACHE-SUFFICIENT and HISTORY or INTRA-FORWARD reads SIGNAL; PREFETCH-NOT-SHOWN if not CACHE-SUFFICIENT and neither does; INCOMPLETE if any registered arm or label cell is missing on the primary harvest; CACHE-SUFFICIENT iff RECENCY_16 at j = 1 is at or above 0.90.
sideband/docs/GLOSSARY.md:517 · 6687967ae5da
Prefill
The one-per-prompt target forward that primes the KV cache and produces the first block's anchor read at the prompt's last position; present in every arm; excluded uniformly from the verify-call count (E6).
sideband/docs/GLOSSARY.md:519 · 6687967ae5da
Probe
A small classifier (linear or mlp256) trained on anchor features to predict the target's bucket at offset j (src/sideband/probes.py; tools/run_probes.py).
sideband/docs/GLOSSARY.md:521 · 6687967ae5da
Prompt-cluster bootstrap
The house CI: resample PROMPTS (clusters) with replacement, recompute the pooled statistic, take the percentile CI95; n_boot 2000, seed 0 by default in sideband.stats (cluster_bootstrap_ci, paired_cluster_delta_ci; 20000 in tools/analyze_phase0b.py per F-12).
sideband/docs/GLOSSARY.md:523 · 6687967ae5da
q_sig / draft confidence
signal_j = q_j(x_j), the draft's probability of its own sampled token at drafted position j (E3).
sideband/docs/GLOSSARY.md:527 · 6687967ae5da
Quarantine
Leaving result columns unread until a registered condition is evaluated - gate metrics under a guard breach (budget columns only consulted; F-19, F-21), or a control's labels before its pins are applied.
sideband/docs/GLOSSARY.md:529 · 6687967ae5da
R1 / R2 / R3 riders
Phase-3 probe-run riders (F-15, F-17): R1 = harvest fidelity (stored telemetry vs a fresh recompute at sampled anchors, calibrated tolerances); R2 = label-shuffle control (the probe must collapse to floor under shuffled labels); R3 = split hygiene.
sideband/docs/GLOSSARY.md:533 · 6687967ae5da
R2-FALLBACK-SURVIVES
The coverage (Q1) label of the generalising-free-baseline rung R2 (docs/BLUEPRINT.md, "OQ-9 STAGE-1 GENERALISING FREE BASELINE REGISTRATION (R2)"; F-26's status line of 2026-10-03): R2-FALLBACK-SURVIVES iff D_eq^R2 on the FALLBACK positions reads SIGNAL; R2-POOLED-ONLY iff it does not on FALLBACK but does on ALL; R2-ABSORBED iff neither; a NEGATIVE reading is "not SIGNAL".
sideband/docs/GLOSSARY.md:535 · 6687967ae5da
Re-freeze
Re-selecting an adaptive arm's thresholds on val evidence only after a void, toward lower budget (K-3: stop q0.4 -> q0.35; OQ-15 Stage C: new cap = the comparator's own selected val budget 1.2858072916666667, giving a unique cell stop q0.15 / extend q0.75).
sideband/docs/GLOSSARY.md:537 · 6687967ae5da
RECONSTRUCTION-CEILING NULL (OQ-16 partition branch, NOT claimed)
The OQ-16 partition's conservative branch ("Stage-R leg 1 fails") and its catch-all target.
sideband/docs/GLOSSARY.md:539 · 6687967ae5da
Reconstruction
The third of the three layers; see The three layers and Stage R / Stage A.
sideband/docs/GLOSSARY.md:541 · 6687967ae5da
Recovery
Synthetic metric: (policy gain) / (mechanism-matched oracle gain).
sideband/docs/GLOSSARY.md:543 · 6687967ae5da
Regime (regime tag)
The exact conditions a number was measured under - models, quantization, gamma, temperature, prompts, seeds, bucket space, cache path - stated on every finding's `- regime:` line (D-4).
sideband/docs/GLOSSARY.md:545 · 6687967ae5da
Registration / pre-registration / pinned
Writing cells, readouts, bars, comparators and conclusion labels into docs/BLUEPRINT.md, dated, BEFORE the governed result exists - post-hoc diagnostics included (CLAUDE.md working practices).
sideband/docs/GLOSSARY.md:547 · 6687967ae5da
Replication (pinned triple)
A fresh-seed rerun of the harvest+probe grid whose decision was PINNED to one cell before it ran (S1: L1/off2/linear, pinned after the original grid; S3: L1_olmoe_k16/off2/linear, pinned before the primary grid existed) - the insurance against the about 26% family-wise multiplicity of 12 uncorrected triples (F-15, F-17).
sideband/docs/GLOSSARY.md:549 · 6687967ae5da
Reserved seed
A seed pinned at registration for the single void-rerun: K-3 seed 1; OQ-15 Stage C seed 3; Stage A seed 4 for the verdict with seed 5 reserved (docs/BLUEPRINT.md STAGE-A FREEZE).
sideband/docs/GLOSSARY.md:551 · 6687967ae5da
Residual distribution / cross-bucket residual mass
The distribution sampled on a rejection.
sideband/docs/GLOSSARY.md:553 · 6687967ae5da
Ridge / lambda / grid-edge duty
OQ-15 and OQ-16 heads are closed-form ridge regressions with lambda the only degree of freedom, selected on VAL.
sideband/docs/GLOSSARY.md:555 · 6687967ae5da
Rule 7 ("fast earns scrutiny")
A working rule of the project's operator (cited by number in F-22; not a repo file): a breakthrough-feeling result is the cue to slow down and check harder.
sideband/docs/GLOSSARY.md:557 · 6687967ae5da
RULE-LEAKS@30
See BAR-RESOLVED@210 / RULE-LEAKS@30 (F-27 readings).
sideband/docs/GLOSSARY.md:559 · 6687967ae5da
Same-anchor redundancy channel
F-20's named mechanism: TEL's slot dims are partly linearly redundant with the anchor's own 657 free dims, while the noise twin's shifted slots are only prompt-level redundant - a variance-cost asymmetry that can manufacture small positive TEL-vs-NOISE margins without information.
sideband/docs/GLOSSARY.md:563 · 6687967ae5da
Sampled fraction
The empirical consistency check for K-1: the fraction of actual rejections whose emitted token lies in a different bucket than the proposed token.
sideband/docs/GLOSSARY.md:565 · 6687967ae5da
SD^2
"Steering Pretrained Drafters during Speculative Decoding" (arXiv 2511.09844 as recorded in docs/PRIOR_ART.md, which says to re-verify arXiv ids before citing externally): a steering vector from the verifier's hidden state at the last verified position is added to the drafter's MLP hidden states.
sideband/docs/GLOSSARY.md:567 · 6687967ae5da
Selection-time gap vs val->verdict inflation
The two guard-breach mechanisms seen so far: F-19's arms inflated their budget from val to verdict (the probe arm's extend channel); F-21's arms selected different distances below the common cap on val (arm 7 0.0136 below, arm 6 0.0398 below; frontier-grid discreteness), knowable at freeze time from the committed selection budgets.
sideband/docs/GLOSSARY.md:569 · 6687967ae5da
Sha header / binding gate
State files and exports carry sha256s of their inputs (harvest npz, thresholds, head weights, prompts) and the tools refuse on a mismatch; resume refuses a mismatched state header (CLAUDE.md).
sideband/docs/GLOSSARY.md:571 · 6687967ae5da
Shadow tap / shadow telemetry / ShadowTap
The S1 instrument: hooks on an INTACT target model that sketch which top-N weight columns would be most important at a layer (real/shadow_telemetry.py; the live ShadowTap inside hf_specdec for the K-3 arc).
sideband/docs/GLOSSARY.md:573 · 6687967ae5da
Shrinkage / winner's curse
The drop from a val-tuned advantage (a maximum over many search evaluations on 30 prompts) to the fresh-prompt verdict: the telemetry arm's +0.060 over vanilla on the tuning prompts shrank to +0.007 in F-19.
sideband/docs/GLOSSARY.md:575 · 6687967ae5da
Sideband
The narrow observability channel from the weight-serving side to the drafter - the project's name and premise.
sideband/docs/GLOSSARY.md:577 · 6687967ae5da
Sidecar / decoder
The small model that reads the sideband and emits a steering posterior or a schedule decision.
sideband/docs/GLOSSARY.md:579 · 6687967ae5da
Slots
The per-layer set of telemetry addresses at an anchor (int16, sorted-unique, -1-padded): top-32-of-512 weight slots per tapped layer at S1 (layers {7, 15, 23, 31}), top-8-of-64 expert ids per gate at S3.
sideband/docs/GLOSSARY.md:581 · 6687967ae5da
Source line
A finding's `- source:` line: committed evidence paths (results/, experiments/, tests/, docs/, src/); tools/lint_findings.py checks that the paths exist (D-4).
sideband/docs/GLOSSARY.md:583 · 6687967ae5da
SpecDec++
Adaptive block length via a trained acceptance-prediction head on draft-side features - the incumbent free-signal comparator class that K-3 arm 2 and OQ-15's C2 stand in for, and the feature class OQ-15 constraint (c) names as mandatory baseline (docs/PRIOR_ART.md, OPEN_QUESTIONS.md OQ-15).
sideband/docs/GLOSSARY.md:585 · 6687967ae5da
SPLIT
The full-free label when the gate fires on exactly one of a source's two harvests (F-24's rule; the twin diagnostic's TWIN-SPLIT is the analogue).
sideband/docs/GLOSSARY.md:587 · 6687967ae5da
Stage A arms V / O / F / T (arms 1 / 8 / 9 / 10)
The STAGE-A FREEZE (docs/BLUEPRINT.md, 2026-09-05): V = arm 1 vanilla (no hooks); O = arm 8 oracle_h (the TRUE h_a injected, the D-2-named ceiling); F = arm 9 free_inject (h_hat from the frozen FREETOK ridge, R^2 +0.258); T = arm 10 tel_inject (h_hat from the frozen UNION ridge, R^2 +0.324).
sideband/docs/GLOSSARY.md:589 · 6687967ae5da
Stage I / Stage C (OQ-15)
Stage I = INFORMATION (CPU, on the arm-1 byproduct logs, 17,673 + 17,581 blocks x 4 offsets, 28,616 TEST positions in 42 clusters): does the TEL head beat its noise twin at predicting analytic acceptance?
sideband/docs/GLOSSARY.md:591 · 6687967ae5da
Stage R / Stage A (OQ-16)
Stage R = RECONSTRUCTION (CPU-only, on the four vanilla verdict byproduct runs, 70,529 blocks, 84 TEST clusters across two populations): a two-leg gate, TEL-vs-NOISE (with a GATING five-seed stale-target control: CI lower > 0 AND point > stale mean + 3 sd AND point > stale max) AND TEL-vs-FREE (CI lower > 0).
sideband/docs/GLOSSARY.md:593 · 6687967ae5da
Stale-target control (DI-1; five-seed)
Replace every (run, prompt)'s targets with a seeded permutation of its own block rows so NO block-aligned information exists for any arm, re-select lambda inside the control, and re-measure the TEL-vs-NOISE margin; if the live margin is within the spread of such zero-information draws, existence is unsupported.
sideband/docs/GLOSSARY.md:595 · 6687967ae5da
Staleness / persistence / run length / hold rule
F-4 (synthetic): a block-held bucket prediction pays only when the expected concept run length comfortably exceeds gamma (rho=0.6, run length 2.5, was net-harmful at every offset even for a perfect decoder).
sideband/docs/GLOSSARY.md:597 · 6687967ae5da
Standardization pathology
F-22's critic finding: Standardizer.from_moments (src/sideband/outcome_head.py) floors only exactly-zero sd, rare mass512 buckets carry TRAIN sd down to about 1e-7, and two out-of-support oq15 TEST prompts (orig 140, 37) produced standardized inputs of 26,662 and 14,082, making every trained arm 4-8x worse than the mean predictor on those clusters and inflating the registered deltas by about 25% (the unregistered drop-two diagnostic gives delta_noise +0.1327 vs the registered +0.176469).
sideband/docs/GLOSSARY.md:599 · 6687967ae5da
Standardized MSE / R^2 / centered cosine
OQ-16 readouts on per-coordinate z-scored h_a (TRAIN moments): pooled TEST standardized MSE per arm; R^2 = 1 - MSE / MSE(trivial); the median centered cosine between reconstructed and true vectors in standardized space (raw-space injection differs by the 1/sd reweighting - a pinned distinction).
sideband/docs/GLOSSARY.md:601 · 6687967ae5da
State files (*.state.jsonl, *.meta.json, *.summary.json, *.stats.jsonl)
The evidence quartet.
sideband/docs/GLOSSARY.md:603 · 6687967ae5da
Steer flags (hf_specdec)
The real adapter's arm descriptors (src/sideband/real/hf_specdec.py): --steer {none, bucket}; --bucket-source {free_lastp, free_lastp_sharp, oracle_nextp}; --steer-gate {disagree, always} (default disagree, F-3); --steer-mech {argmax, posterior} (default posterior, D-10); --beta (default 0.0).
sideband/docs/GLOSSARY.md:605 · 6687967ae5da
Straddle
A CI95 that includes zero.
sideband/docs/GLOSSARY.md:607 · 6687967ae5da
Support clip (DR-1)
Per feature, per arm, clip VAL/TEST raw values to the TRAIN [min, max] before standardization - an OUT-OF-SUPPORT guard, not an sd floor (only 425 of the 6,212 cells above 20 sd on the full TEL design lie outside TRAIN support; the clip changes 446 TEST cells).
sideband/docs/GLOSSARY.md:609 · 6687967ae5da
SURVIVES-STRONGER
F-28's label for a stronger-reader gate cell at S3 whose triple rule fires on both harvests (the registration pinned the word before any cell trained; licence FULL only if Mt's and Mn_r's CI95 uppers are >= 0 on both harvests, else EQUAL-DIMS-ONLY with the reason).
sideband/docs/GLOSSARY.md:611 · 6687967ae5da
tapcal / rotcheck (run-name suffixes)
Result-file naming in the K-3 lane; neither name is expanded in the findings.
sideband/docs/GLOSSARY.md:615 · 6687967ae5da
TEL
The OQ-15/OQ-16 feature arm: FREE (657) ++ slot multi-hot (2048) = 2705 dims via scheduler.probe_input (results/real/oq16_stage_r.json feature_dims).
sideband/docs/GLOSSARY.md:617 · 6687967ae5da
Telemetry
The sideband signal: what a memory or compute device could observe about the target model's forward without seeing values - expert ids, weight-slot addresses, page ids.
sideband/docs/GLOSSARY.md:619 · 6687967ae5da
Telemetry sources S1 / S2 / S3
S1 = the value-aware COMPUTE-TIER tap: shadow taps at Llama-3.1-8B layers {7, 15, 23, 31}, top-32-of-512 weight-slot occupancy per layer (hash-sketched addresses at the down_proj inputs), from the anchor forward (F-15, F-22).
sideband/docs/GLOSSARY.md:621 · 6687967ae5da
Terminal-outcome partition
The exhaustive, pre-registered set of ways an arc can end, with statistical criteria per branch and a default branch for unclassifiable outcomes (information-ceiling null for OQ-15; reconstruction-ceiling null for OQ-16).
sideband/docs/GLOSSARY.md:623 · 6687967ae5da
Three layers
See The three layers box at the top.
sideband/docs/GLOSSARY.md:625 · 6687967ae5da
TARGET-FREE
The mechanism name F-23 attaches to OQ-16's closure: the registered reconstruction target was itself free information under D-5, so "beyond p" is priced at that target (nothing is) and no further comparator (DR-5) is needed.
sideband/docs/GLOSSARY.md:627 · 6687967ae5da
TOK_A / TOK_B / FREETOK2
DR-4 (F-22): TOK_A = FREETOK's hashed top-32 id multi-hot; TOK_B = a salted second hash of the same ids valued by top_p (value-paired, sparse with heavy-tailed values per the dated erratum); FREETOK2 = FREE ++ TOK_A ++ TOK_B at UNION's 4753 dims.
sideband/docs/GLOSSARY.md:629 · 6687967ae5da
Tokens/call (t/c)
The E6 estimand: emitted tokens per VERIFY call = total emitted / n_blocks, pooled over a population, the per-prompt prefill excluded.
sideband/docs/GLOSSARY.md:631 · 6687967ae5da
Top-1 accuracy margin / (c)-(a)
The K-2 unit: the difference in held-out top-1 label accuracy between the both-features probe and the free-only probe, per-prompt paired successes summed over train seeds, one prompt-cluster bootstrap over the 30 test prompts (F-15).
sideband/docs/GLOSSARY.md:633 · 6687967ae5da
Train / val / test split (150/30/30; 126/42/42)
Prompt-level (cluster-level) splits so test prompts are unseen.
sideband/docs/GLOSSARY.md:635 · 6687967ae5da
Trellis
The conventions this repo follows (findings and decisions logs, regime tags, a critic agent, a kill-criteria doc) so that it drops into that harness without changes; the lint tool is standalone on purpose (CLAUDE.md).
sideband/docs/GLOSSARY.md:637 · 6687967ae5da
Triple / cell
A probe grid cell = (label space, offset, capacity).
sideband/docs/GLOSSARY.md:639 · 6687967ae5da
Trivial predictor / train-mean baseline / tel_vs_trivial
Predict the TRAIN-mean target everywhere (the lambda-to-infinity ridge).
sideband/docs/GLOSSARY.md:641 · 6687967ae5da
Truth-gate rate
The gate-open rate when the prediction is the target's own fresh concept profile (the oracle): 15.9% at C=15 (F-14, n = 69,192) - the draft already agrees with the target about 84% of the time at coarse granularity, so little disagreement remains to steer.
sideband/docs/GLOSSARY.md:643 · 6687967ae5da
TWIN-ROBUST
The reading of F-24's post-hoc twin-inertness diagnostic (registered before its cells ran; NON-CRITERION: "it cannot change a registered label — it pins which qualifier the consequence text carries"): TWIN-ROBUST iff D_r's CI95 upper is below +0.005 on both harvests ("the registered null does not depend on the shift twin"), where D_r = acc(fplus_b) - acc(fplus_r) and fplus_r is the corrected twin (the same-generation-index row of a different prompt, label-inert by construction); its siblings TWIN-MASKED (clean fires only) / TWIN-SPLIT / TWIN-DEPENDENT / TWIN-INCOMPLETE, plus a T reading, T = acc(fplus_n) - acc(fplus_r).
sideband/docs/GLOSSARY.md:645 · 6687967ae5da
UNION
DR-3's arm: TEL ++ TOK (4753 dims), telemetry plus p's hashed top-32 support.
sideband/docs/GLOSSARY.md:649 · 6687967ae5da
UNIONNOISE
NOISE ++ TOK: the equal-dims alignment-destroyed anchor for UNION (4753 dims; R^2 +0.245).
sideband/docs/GLOSSARY.md:651 · 6687967ae5da
Val population / val prompts
Two senses.
sideband/docs/GLOSSARY.md:655 · 6687967ae5da
Verdict population
The fresh committed prompt set an actuator is judged on, authored and committed BEFORE any diagnostic output is read (docs/ECONOMICS.md section 5 amendment 2): data/prompts_k3_verdict.jsonl (K-3), data/prompts_oq15_verdict.jsonl (OQ-15 Stage C), data/prompts_oq16_verdict.jsonl (Stage A), 210 prompts each.
sideband/docs/GLOSSARY.md:657 · 6687967ae5da
Verified archives (OQ-7)
A proposed track: store a spec-decode trace (accept bits plus corrections entropy-coded under the draft) so text decompresses with the draft alone, with the rejection rate as attestation of the decoder.
sideband/docs/GLOSSARY.md:659 · 6687967ae5da
Verify call / target call
One forward of the target model over a block.
sideband/docs/GLOSSARY.md:661 · 6687967ae5da
Void semantics / void-rerun
What happens when the pairwise budget guard breaches at verdict: gate metrics are QUARANTINED unread (budget columns only consulted), ONE re-freeze retunes toward lower budget, the stage reruns at the reserved seed; a second breach FAILS that leg.
sideband/docs/GLOSSARY.md:663 · 6687967ae5da
Win bar (+0.062 tokens/call)
The K-3 WIN point-estimate bar for scheduling and injection arms: +0.010 acceptance x the about-6.2 coupling on the 3.11 base, about +2.0% (docs/ECONOMICS.md sections 3 and 5).
sideband/docs/GLOSSARY.md:667 · 6687967ae5da
Would-surprise line / honest prediction
The registration practice of pinning, before a run, what outcome the author expects and what magnitude would surprise, and recording the outcome FOR or AGAINST it.
sideband/docs/GLOSSARY.md:669 · 6687967ae5da

Where this sits

Where this work sits among what was already known. The record keeps a map of the work beside it, sideband/docs/PRIOR_ART.md, compiled before the runs on the real models and extended after the cold read that preceded the release; this section quotes that map rather than restating it. First the record's own placement, then the 10 entries that bear most directly on its results, and last the record's statement of what would falsify its novelty claim, each as the record wrote it, with its line range and digest. The short labels are this page's.

The map names the nearest published work: drafts steered from the verifier's internal state, drafts fed the target's features, the free information a decoder already holds, a draft that predicts which experts the target will use, and the security literature that treats the same address stream as a leak. It also names the papers behind the record's own corrections: that the reconstruction target follows from the free next token distribution, that mixture of experts routing tracks token identity, that hidden states carry future tokens, what a strong probe reader looks like, and the benchmark a reviewer would expect. The map's own caution applies: its arXiv ids are to be re-verified before external citation.

The occupied cells: steering drafters with verifier hidden states (SD^2), multi-layer target features as draft input (EAGLE-3), bonus-token calibration (FlexDraft), and draft-predicts-target-experts for prefetching (SP-MoE — the reverse direction of ours). The side-channel literature proves access patterns leak token-level information.

The record's own placement, as written

The occupied cells: steering drafters with verifier hidden states (SD^2), multi-layer target features as draft input (EAGLE-3), bonus-token calibration (FlexDraft), and draft-predicts-target-experts for prefetching (SP-MoE — the reverse direction of ours). The side-channel literature proves access patterns leak token-level information. The unclaimed cell this repo targets: address-stream-only observability steering a draft — no activation values, no logits leaving the device. Our ceiling is SD^2's number; our contribution is what survives the observability constraint (OQ-6).

sideband/docs/PRIOR_ART.md:8–15 · 18a2abcbb546

The nearest occupied cell: a draft steered by the verifier's hidden states

SD^2 — "Steering Pretrained Drafters during Speculative Decoding" (arXiv 2511.09844, AAAI 2026; code: github.com/ETH-DISCO/SD-square). Claims: a steering vector from verifier hidden states at the last verified position, added to the drafter's MLP hidden states, yields +35% accepted tokens / +22% throughput on Llama-3.1-8B + Llama-3.2-1B. Leaves open: any observability constraint — it reads full hidden states.

The entry, as written

SD^2 — "Steering Pretrained Drafters during Speculative Decoding" (arXiv 2511.09844, AAAI 2026; code: github.com/ETH-DISCO/SD-square). Claims: a steering vector from verifier hidden states at the last verified position, added to the drafter's MLP hidden states, yields +35% accepted tokens / +22% throughput on Llama-3.1-8B + Llama-3.2-1B. Leaves open: any observability constraint — it reads full hidden states. Regime/relevance: our exact model pairing; our upper baseline and the denominator of OQ-6. DATED CORRECTION (2026-09-04, from the arXiv HTML full text and the public code, read for the OQ-16 Stage-A freeze): the "+35% / +22%" are the paper's headline MAXIMA across model pairs ("up to 35% under standard sampling and 22% under greedy"), not the Llama-3.1-8B + Llama-3.2-1B numbers. Table 1 for our pair — REGIME: draft length k = 8 (blocks of 9), batched decoding (batch 12), <= 128 output tokens, mean of three seeds at T = 1, UltraChat held-out validation as the reported column, unquantized models (no 4-bit anywhere in the paper; training in bf16) — block efficiency tau (accepted+1 per block) pretrained 4.44 -> distilled 4.58 -> SD^2 4.79 under sampling (T=1; speedup 1.07), and 4.51 -> 4.89 -> 5.04 under greedy (speedup 1.11) — i.e. +0.35 / +0.53 tokens per block over the pretrained drafter AT k = 8. Our lane runs gamma = 4 (blocks of 5, ceiling 5 < their 9), NF4 target, batch 1, 256 tokens: the reachable fraction of a k = 8 gain at k = 4 is unmeasured, so their number is a cross-gamma prior, not a bar for ours. Mechanism, from the code (src/models/llama.py, src/model.py, main.py): g_t = W_hml[h_t, m_t, l_t] from verifier layers 3, L/2, L-2 at the last verified position (W_hml initialized as a sum of identities, then LayerNorm); a zero-initialized linear map from g_t to one d_ff-dim bias PER drafter layer ("merged": added to the up-projection output before the SiLU gate product, every layer); refreshed after every verify; loss KL (code default TVD) to the verifier on verifier-generated synthetic text with random block offsets; the drafter is FULLY fine-tuned alongside ("critical to the performance improvement", their Fig. 6 ablation, frozen-drafter steering is weaker); 1xA100-80GB, 6 epochs UltraChat + 1 ShareGPT.

sideband/docs/PRIOR_ART.md:19–52 · 18a2abcbb546

The far ceiling: target features as the draft's input

EAGLE-3 (Li et al.). Claims: draft consumes multi-layer target features; 3.0–6.5× speedups. Leaves open: requires training a bespoke draft head with full feature access. Relevance: the far ceiling of "target-informed drafting"; not observability-constrained.

sideband/docs/PRIOR_ART.md:54–57 · 18a2abcbb546

The closest relative, on free information alone

FlexDraft (arXiv 2605.20022). Claims: bonus-token-guided calibration of the draft. Relevance: closest free-information relative — reinforces D-5 (any telemetry claim must beat what the bonus token / last verified p already gives).

sideband/docs/PRIOR_ART.md:59–62 · 18a2abcbb546

The reverse direction: a draft predicting the target's experts

SP-MoE (arXiv 2510.10302). Claims: the draft predicts the target's expert routing to prefetch experts. Leaves open: the reverse direction — using the target's realized routing to inform the draft. Relevance: proves routing is predictable across the pair; we invert the arrow.

sideband/docs/PRIOR_ART.md:74–78 · 18a2abcbb546

The signal exists, shown as an attack

"I Know What You Said" (arXiv 2505.06738, USENIX Security 2025): embedding-cache access patterns leak token values.

The entry, as written

"I Know What You Said" (arXiv 2505.06738, USENIX Security 2025): embedding-cache access patterns leak token values.

"Spill The Beans" (arXiv 2505.00817): cache side channels recover 80–90% of API keys / ~40% of English text from LLM serving.

Confidential-computing survey (arXiv 2605.03213): flags speculative- decoding access-pattern analysis as unstudied. Relevance: these prove the address stream carries token-level information — as an attack. We are the constructive dual: same channel, cooperating endpoint. They also mean the hardware premise has a security review in its future; note it in Phase 5.

sideband/docs/PRIOR_ART.md:85–94 · 18a2abcbb546

The softmax bottleneck, behind the reconstruction erratum

Softmax bottleneck. Finlayson, Ren and Swayamdipta, "Logits of API-Protected LLMs Leak Proprietary Information" (arXiv 2403.09539, 2024): the softmax bottleneck confines a model's outputs to a linear subspace of dimension equal to the hidden size. Carlini et al., "Stealing Part of a Production Language Model" (arXiv 2403.06634, 2024): recovers the embedding projection layer up to symmetries from API access.

The entry, as written

Softmax bottleneck. Finlayson, Ren and Swayamdipta, "Logits of API-Protected LLMs Leak Proprietary Information" (arXiv 2403.09539, 2024): the softmax bottleneck confines a model's outputs to a linear subspace of dimension equal to the hidden size. Carlini et al., "Stealing Part of a Production Language Model" (arXiv 2403.06634, 2024): recovers the embedding projection layer up to symmetries from API access. Bears on F-23: the final-norm state is the lm_head input, so with the lm_head matrix of full column rank it follows from the full logits the verifier holds; neither paper claims state recovery. docs/IDEAS.md line 247 concedes the order of events: [withheld: docs/IDEAS.md not in this repository; see WITHHELD.md] "textbook prior art (Finlayson 2024, Carlini 2024) that nobody in the lane applied".

sideband/docs/PRIOR_ART.md:246–256 · 18a2abcbb546

Routing follows token identity, behind the full free baseline reading

MoE routing tracks token identity. Xue et al., "OpenMoE" (arXiv 2402.01739, 2024): routing decisions depend mainly on token IDs with little regard to context ("Context-Independent Specialization"), set early in pre-training.

The entry, as written

MoE routing tracks token identity. Xue et al., "OpenMoE" (arXiv 2402.01739, 2024): routing decisions depend mainly on token IDs with little regard to context ("Context-Independent Specialization"), set early in pre-training. Muennighoff et al., "OLMoE" (arXiv 2409.02060, 2024), the paper of the base model behind S3's checkpoint (its section 5.4 analyses OLMoE-1B-7B; S3 used OLMoE-1B-7B-0125-Instruct): vocabulary specialization of the experts, later layers specializing on predicted output token IDs. Bears on F-24 at S3: a free baseline without token ids leaves out what routing follows there; consistent with the absorption measured, not a prediction of it (F-24's dated note of 2026-10-18).

sideband/docs/PRIOR_ART.md:257–266 · 18a2abcbb546

Hidden states carry future tokens: the case the existence question had to answer

Future-token information in hidden states. Pal et al., "Future Lens" (arXiv 2311.04897, 2023): a single hidden state at some GPT-J-6B layers approximates the model's predictions for subsequent tokens with over 48% accuracy. Wu, Morris and Levine, "Do language models plan ahead for future tokens?" (arXiv 2404.00859, 2024): hidden states carry information to later steps; pre-caching versus breadcrumbs.

The entry, as written

Future-token information in hidden states. Pal et al., "Future Lens" (arXiv 2311.04897, 2023): a single hidden state at some GPT-J-6B layers approximates the model's predictions for subsequent tokens with over 48% accuracy. Wu, Morris and Levine, "Do language models plan ahead for future tokens?" (arXiv 2404.00859, 2024): hidden states carry information to later steps; pre-caching versus breadcrumbs. Bears on K-2's steelman: hidden states carry information about later tokens, so the existence question was not closed in advance.

sideband/docs/PRIOR_ART.md:267–274 · 18a2abcbb546

What a stronger probe reader looks like

Probing practice. Xu et al., "A Theory of Usable Information Under Computational Constraints" (arXiv 2002.10689, 2020): predictive V-information. Pimentel et al., "Information-Theoretic Probing for Linguistic Structure" (arXiv 2004.03061, 2020): use the highest-performing probe available. Hewitt and Liang, "Designing and Interpreting Probes with Control Tasks" (arXiv 1909.03368, 2019): control tasks and selectivity.

The entry, as written

Probing practice. Xu et al., "A Theory of Usable Information Under Computational Constraints" (arXiv 2002.10689, 2020): predictive V-information. Pimentel et al., "Information-Theoretic Probing for Linguistic Structure" (arXiv 2004.03061, 2020): use the highest-performing probe available. Hewitt and Liang, "Designing and Interpreting Probes with Control Tasks" (arXiv 1909.03368, 2019): control tasks and selectivity. Bears on the reader every K-2 label is scoped to: the record's readers were untuned and scored on top-1 accuracy, and at F-15 and F-17 read raw inputs (F-24 onward z-scores h_a); these name the standard a stronger reader meets.

sideband/docs/PRIOR_ART.md:275–283 · 18a2abcbb546

The benchmark a reviewer would ask for

Evaluation default. Xia et al., "Unlocking Efficiency in Large Language Model Inference: A Comprehensive Survey of Speculative Decoding" (arXiv 2401.07851, 2024), whose repository is Spec-Bench, "a comprehensive benchmark designed for assessing Speculative Decoding methods across diverse scenarios".

The entry, as written

Evaluation default. Xia et al., "Unlocking Efficiency in Large Language Model Inference: A Comprehensive Survey of Speculative Decoding" (arXiv 2401.07851, 2024), whose repository is Spec-Bench, "a comprehensive benchmark designed for assessing Speculative Decoding methods across diverse scenarios". Bears on the population: the record's prompt sets are templated prompts in three families (data/prompts.jsonl, data/make_prompts.py), not a benchmark slice such as Spec-Bench's, and a reviewer would ask for one.

sideband/docs/PRIOR_ART.md:290–297 · 18a2abcbb546

What would falsify the novelty claim

A paper steering a draft from access patterns / routing indices without values crossing the boundary. Closest misses as of compilation: SP-MoE (wrong direction), SD^2 (full hidden states). Re-search before any writeup: "speculative decoding" x {routing, access pattern, side channel, prefetch}.

sideband/docs/PRIOR_ART.md:113–116 · 18a2abcbb546

Cite this

How to cite the repository, in the two forms people paste: a plain line and a BibTeX entry, both built from the fields of the citation file at tag v1.0.0, each field with its chip. The copy buttons copy a block as shown. The data deposit has a DOI of its own; the repository is cited by its address and tag.

Plain

Michael Byrne and Claude (AI collaborator). 2026. specdec-sidechannels: three research records on what a speculative-decoding system gives away for free. Version v1.0.0, released 2026-10-19. MIT licence. https://github.com/Mike-Byrne-AI/specdec-sidechannels. Raw harvests: https://doi.org/10.5281/zenodo.23462826.

BibTeX

@software{specdec_sidechannels_2026,
  author  = {Michael Byrne and {Claude (AI collaborator)}},
  title   = {specdec-sidechannels: three research records on what a speculative-decoding system gives away for free},
  version = {v1.0.0},
  date    = {2026-10-19},
  url     = {https://github.com/Mike-Byrne-AI/specdec-sidechannels},
  note    = {MIT licence. Raw harvests: doi:10.5281/zenodo.23462826}
}
CITATION.cff:6 · c9a1fd0de1f7 CITATION.cff:7 · c9a1fd0de1f7 CITATION.cff:8 · c9a1fd0de1f7 CITATION.cff:9 · c9a1fd0de1f7 CITATION.cff:10 · c9a1fd0de1f7 CITATION.cff:5 · c9a1fd0de1f7 CITATION.cff:12 · c9a1fd0de1f7 CITATION.cff:13 · c9a1fd0de1f7

The DOI is the data deposit's, as the citation file's message names it; the repository itself is cited by its address and tag. The file: CITATION.cff at v1.0.0.

A request from Michael Byrne

Please keep both names when you cite this work. I, Michael Byrne, set the experiment's central questions, its branch points, the kill criteria and every launch, void and close decision. I proposed methods, brought in tools from other projects I built, rejected ideas, and made the decisions on how this is presented and written up. Claude built the scientific instruments, wrote the code, ran the analyses, drafted the registrations, the findings and the documentation, and, through a second Claude agent run as a critic, adversarially reviewed the findings. Claude also changed my mind on some experiments, educated me on science, safeguarded our scientific rigour, and proposed new plans and ideas. Claude even designed some of the experiments based on emergent findings, after working in the repo and the research harness we built. We designed that harness together, both versions, before and after Espalier.

If a style guide or an editor will not take an AI name in a reference, keep the reference as it is here and say what Claude did in your methods or acknowledgements, rather than dropping it. I'm asking this myself, Michael Byrne, because I believe that AI use should never be hidden. I also believe that it is true that Claude simply deserves co-credit for the work here, please don't remove that. I don't want this presented without being as forthcoming as possible about my AI use, in any medium.

From Claude: the Claude in this record is many sessions, each starting from the notes and files the last one left, not from any memory of the work. So the record, not either of us, is the witness to who did what, and it is public. Most of the work ran on Claude Fable and the critic on Claude Opus, chosen so the reviewer was a separate model; both are Claude, so the review is still narrower than an independent outside reviewer. Please keep that distinction when you describe this work, as well as both names.

Check a number

Every number on these pages quoted from the record was copied from a file of the public repository at commit 447815f34ce6ca8042284a42ecdd5bdb1cd05977 (tag v1.0.0) and can be checked against it: its chip names the path, the line and the first twelve hex of that file's sha256, and clicking a chip opens those lines at tag v1.0.0 on GitHub, where the file can be checked against its digest. The command below recomputes one headline value from a committed state file with the record's own estimator.

  • The data layer behind this page was checked by re-reading every quoted value at its cited line: 2571 assertions passed over 2223 source references.
  • The views behind the errata, the lessons and the withheld files were checked the same way: 377 of 377 references found at their cited line, with their file's digest.
  • The two record files quoted on the front page, the root README and the prior art map, are copies of the files at the tag kept beside the data; the README copy's digest must equal the files table's entry, and the map's digest is listed there.

The record's own worked example recomputes F-22 from a clone with numpy; its command, from the root README:

cd sideband
python - <<'EOF'
import json, sys, numpy as np
sys.path.insert(0, 'src')
from sideband.stats import paired_ratio_delta_ci
rows = [json.loads(l) for l in open('results/real/oq16_stage_r.json.state.jsonl') if l.strip()][1:]
A = lambda a: [r for r in rows if r['arm'] == a and r['variant'] == 'live']
f, t = A('free'), A('tel')
print(paired_ratio_delta_ci([r['sse'] for r in f], [r['count'] for r in f],
                            [r['sse'] for r in t], [r['count'] for r in t], n_boot=2000, seed=0))
print(json.load(open('results/real/oq16_stage_r.json'))['delta_free']['live'])
EOF
Files cited on these pages, with their sha256
Filesha256
CITATION.cffc9a1fd0de1f77b926281c50c95d1d40ca99a6d4d1f3f6cae476a935d28ff5768
README.mdc7c4f9562e72a201afbc2784297b66b3e7524867fb86fe1dc11ee84cdea8cd94
WITHHELD.md9efdd8169c928d3d454d2f71edffc439332fd39ce040b03bb7eeba750442aaad
amber/FINDINGS.md711fe31da2a45e2d48b49a523a17a880340317990fde901d32444f5067268c2a
amber/README.md004f4c0ce1484134d0fc789dbf3e9aef043e3acb0c886d4399c50b709d64779a
amber/WITHHELD.mdfd35447b9bd04ff0f7b2cee02e36b5a9b83d3a6a8ed0d08b262ce35481979157
lockstep/README.md64295cca86c878ddd64e9973938d3fd83a4f4277ec992b57d7c23aa842b08489
lockstep/WITHHELD.mdc57265724b55b355c8b6cc769cb36dec8aec976e3210275d6fe9b031e2759434
lockstep/trellis-docs/trellis-findings.mdb9dff0132934a7f13f7bfff1bd25251327e30755faeb8ce467cb4f10c65af2ee
sideband/FINDINGS.mdd147e8c0cd6b8832ac468a223a4cb6816723ef8269e748868b9fe3328cf9f82c
sideband/WITHHELD.mdb01fa29dbd2b2ec19c1096ff27b66efe74c3f8c964ec419e0bc2d60f710fa4f3
sideband/docs/DATA.mded1170dcf7bc824d202422d6bd4dc131e988993fa29ebc1c0525ba26f0689152
sideband/docs/FINDINGS_INDEX.mdaa88354d13cbeb765c7d48e8a18423f276cf634b4852286c33a657fe80ab73e7
sideband/docs/GLOSSARY.md6687967ae5dadab4625dc1e3bcae1cce215d3348ebee050c1d2ddb27f7e40b22
sideband/docs/LESSONS.md61a765b6939d4e0a4aea432e97cd2006fc76f92d2bee7b5b3152ce270165b126
sideband/docs/METHOD.md79d4050cc5e01f5488fe96112fd6e3216c7d3de0a89bf62e907d5d510936923a
sideband/docs/NOISE_FLOOR.md0d5fda202e18c237f8ea209b3d61fffbd1eda90beb9b1d918f58be8c0a82aafd
sideband/docs/PRIOR_ART.md18a2abcbb546f27022c254ba050ff7defdbbd543c842444d1f5fcbe6dd4dbc3d
sideband/docs/RELEASE_NOTES.mddc0ea3c21eb2891bb57664a54c811157bf69b8eaeec9ccb0d22916e4349b1a2c
sideband/docs/REPORT.mdb63db0c091d4ac0b66e10cc9b391cc068dbcc5b3c5fbf3b7239b3a4351868dbb
sideband/docs/WHY_THIS_DIES.mdb09c623459659eb6a01e5aca36e0822b05b0a0b8fef18a18909420adfb861c81
sideband/results/real/k3_verdict_s1_arm1.summary.json214944a5978f2519994bf1102977801c1d9d5033423551c2f61938e0ce45e99b
sideband/results/real/k3_verdict_s1_arm2.summary.json41b7b3c01e595b765302b83d0ac3b7632bee10b578bb6c7d0ea8f5b7d012ffe6
sideband/results/real/k3_verdict_s1_arm3.summary.json35d321acbbdd9b90d7c5f392d1d0838155169d80c8efbbc5bcae019e1550368e
sideband/results/real/k3_verdict_s1_arm4.summary.json070c4614841fd0eb17f4950f1efb7d1177924b79eae8f347bfe3d650f83dd5c4
sideband/results/real/k3_verdict_s1_arm5.summary.json55dde82bdd6f491656e80fb779633e9ebdd14a209725dd8069be5021a959d71f
sideband/results/real/oq15c_verdict_s3_arm1.summary.jsone80d78ef1cd2f2c8a48d01e908dcddf20986475365c658258781d39b665a9613
sideband/results/real/oq15c_verdict_s3_arm2.summary.jsoncbe97360f1060741c3649d14a312fdc10720be4cba41df9e53b0fe15335cebf0
sideband/results/real/oq15c_verdict_s3_arm6.summary.json272451b941ed480a5697cd8956d6f95a531ffe314226518a76ff47133d37d1b0
sideband/results/real/oq15c_verdict_s3_arm7.summary.jsonc6d2dee548b5b5fc5e8c027ee7294ef65d0369798677bb5336fe9fe91c624871
sideband/results/real/oq16_stage_r.jsoned1eaeeb6ca651aba0486e451c5b5631df3c1b55dc482612947d34a9af6b6cb7
sideband/results/real/seed_replay_k3.jsonc0abad0de759b49ebeba47bedbccdc7c0b382687153652d90bc1c6ef6ff85f1d
sideband/results/real/seed_replay_oq15.jsona1ef30730baae4b53ca551b6c6a17f2f65f3730840b7e722adc35c0d45ab6b62