What a speculative decoder gives away for free

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)

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)

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)

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)

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)

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)

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)

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)

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

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)

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

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

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

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

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]

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)

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

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

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

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]

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

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]

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

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; …

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]

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; …

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)

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; …

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

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]

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; …

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

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]

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); …

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)

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; …

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

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

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

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]

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

`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

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

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

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

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

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

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

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

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

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

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

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