B4C02 — Corrected Full-N Result Report

Gate: H-AIF-G5 · Cell: B4C02_MISSPECIFIED_WORLDS · Lane: B · Date: 2026-07-22 Result: hierarchical-aif/results/motor_stack_aif/B4C02_CORRECTED_FULL_RESULT.json sha256 0633988dbfd690c0c0d12075dba4e0d8c25ddd178125064bd01fbdaf4629e398 Prediction record: hierarchical-aif/protocols/B4C02-CORRECTED-FULL-PREDICTION.md Maps to existing ladder: constrains the interpretation of P3. No P-level is raised.

1. Executive result

B4C02 ran to completion at the full frozen N = 200 simulations per generator, 600 total, with zero failures and resourceBoundPartial: false. runStatus = ELIGIBLE_FOR_FROZEN_VERDICT.

Observed gensWithM2overM3 = 1 of 3 → verdict GENERATOR-SPECIFIC → the frozen prediction GENERATOR-ROBUST_ADVERSE is REFUTED at full N.

By the outcome mapping committed in §6 of the prediction record, this means: the adverse M2-over-M3 result is NOT generic to heavy-tailed dwell shape. H_SHAPE_ARTIFACT is weakened. Per the committed interpretation, this strengthens the case that the M2-vs-M3 contrast carries mechanism-relevant information — but it establishes no mechanism by itself, and this report makes no mechanistic claim.

2. The three generators — full results

gensWithM2overM3 = #{generators with m2_beats_m3_frac >= 0.5} = **1**. Frozen criterion: >= 2 of 3 → GENERATOR-ROBUST_ADVERSE; otherwise GENERATOR-SPECIFIC. 1 ≤ 1 → GENERATOR-SPECIFIC.

Full winner frequencies (200 sims each, 6 competitors)

The separation is extreme, not marginal. M2 wins 187/200 under the three-timescale heavy-tail world and 1/200 under each of the other two. This is not a threshold that was barely missed; two of the three generators put M2's win rate at half a percent.

What the pattern says, stated carefully

The one world in which the lognormal reliably out-predicts the two-timescale mixture is the world whose truth is a three-timescale heavy tail — i.e. a process with more timescales than M3 can express. Under a Weibull–Gamma blend or per-motor heterogeneous Weibull, M3 wins outright.

That is a hypothesis-generating observation about the simulated worlds, not a finding about the real cohort. It is consistent with — and does not establish — the reading that the real data's M2-over-M3 gap reflects unmodelled tail structure rather than generic shape. Distinguishing those requires evidence this cell does not contain.

3. Provenance

Determinism. PYTHONHASHSEED was unset, and that is now irrelevant: the D3 fix replaces hash(gen_label) with seeding.stable_seed, a SHA-256-derived integer. This run is reproducible; a run of the committed cell would not have been, because the committed cell produces different synthetic data on every invocation. This run is therefore not bit-comparable to the committed cell — it could not have been, and the prediction record said so in advance.

D2 note. The recorded RESOURCE_BOUND reason claimed 150–250 h. The measured cost is 8.17 h, against the re-measured estimate of 8.7 h — an 18–31× overstatement in the original, and the first cost estimate this programme has made that came in accurate.

4. Prospectivity — SATISFIED, with the one caveat stated

This is the only cell in the current batch whose prediction was committed before its observation.

Caveat, stated rather than omitted. The commit occurred mid-run. Reconstructing the generator timeline from the recorded runtimes (17:51:02Z start; 11 227 s + 9 266 s + 8 916 s), at the moment of commit weibull_gamma_blend had completed (fraction 0.0050) and three_timescale_heavy_tail was in progress with its progress counter visible in the log. So one generator's outcome, and a trend in a second, were on disk when the prediction was committed.

That cuts in favour of the prediction's honesty rather than against it: with one generator already failing the threshold and one trending above it, the committed prediction GENERATOR-ROBUST_ADVERSE required the third, entirely unobserved generator to pass. It did not. The prediction was committed against the partial evidence available, not tuned to it — and it was refuted.

Graded: PROSPECTIVE_WITH_MID_RUN_COMMIT_CAVEAT. Compare B4C10 and the F-side scoring, both NOT_SATISFIED (D9).

5. Prediction outcome, against the committed mapping

The mapping applied here is the one committed in §6 of the prediction record, not one selected afterwards:

gensWithM2overM3 <= 1 → GENERATOR-SPECIFIC → prediction REFUTED — H_SHAPE_ARTIFACT weakened. "The adverse result is not generic to heavy-tailed shape. This strengthens the case that the M2-vs-M3 contrast carries mechanism-relevant information — but establishes no mechanism by itself."

I predicted GENERATOR-ROBUST_ADVERSE and was wrong. That is the recorded outcome, and being wrong in the pre-committed direction is the point of committing it.

6. Scope boundary — what this cell cannot do, reproduced from §7 of the prediction record

• It cannot establish biological parity, mechanism, or active inference.
• It cannot promote M2 to "the UNI model." M2 is an adversarial baseline.
• It cannot move P6 structural/mechanistic on its own; it constrains the interpretation of P3.
• M3 is not vindicated. M3 winning under two synthetic worlds says nothing about M3 on the real cohort, where the held-out adverse result stands unchanged.
• M4 / M7 / M8 are skipped by construction (the cell's declared design — see D4), so this cell says nothing about the mixture, hierarchical, or KDE models.
• It says nothing about the F-side motor-stack model, which is not entered in this competition.
• These are simulated worlds. No result here is an observation of a bacterium. Nothing in this report may be labelled OBSERVED.

7. The real-data adverse finding is UNCHANGED

The retained B3 adverse result — M2_LOGNORMAL out-predicting M3_TWO_TIMESCALE by ≈0.0369 nats event-pooled on held-out data — stands exactly as before. B4C02 does not overturn it, weaken it, or explain it away. It changes only what may be inferred from it: the pattern is not a generic consequence of heavy-tailed shape, so "it's just shape" is no longer an adequate account.

Independently, the F-side scoring run this session found M2_LOGNORMAL (3.4093) and M8_EMPIRICAL_KDE (3.4225) still out-predicting the F-side motor stack (3.4327) on point estimate, NOT_ESTABLISHED under CI. Both adverse results are reported alongside this one, never instead of it.

8. Lane impact — scoped

9. Limitations

• Simulated worlds only. Three generators do not span the space of plausible misspecification. A fourth world could behave like the third.
• 6 competitors, not 9. M4/M7/M8 are skipped by design, so "M3 wins" means "M3 wins among the six entered."
• Single cohort geometry. Every simulated dataset is built on the derived_eligible_1_to_8 cohort structure — 19 holdout motors — so the sampling noise of that geometry is inherited.
• m2_beats_m3_frac is a win-rate, not an effect size. A 0.94 win rate says how often M2 wins under that world, not by how much, and no interval is attached to the per-generator fractions. The frozen criterion is a count over generators, and that is what is reported.
• Being refuted in the pre-committed direction is a recorded outcome, not a defect.

10. Gate status

Remaining under G5: B4C11 running; B4C01 queued (≈14.5 h measured budget).

NEXT_ACT = commit protocols/B4C11-CORRECTED-FULL-PREDICTION.md BEFORE the running B4C11 result lands (time-limited, requires the human principal — it is the only way B4C11 becomes prospective); continue supervising PID 26756 (B4C11); recompute per-motor NLPD arrays for M4/M6/M7 from their frozen fitted params so they can be contrasted against the F-side candidate under the same paired motor-cluster bootstrap