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

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

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