B4C01 — Corrected Full-N Result Report

Gate: H-AIF-G5 · Cell: B4C01_SYNTHETIC_PARAMETER_RECOVERY · Lane: B · Date: 2026-07-22 Result: hierarchical-aif/results/motor_stack_aif/B4C01_CORRECTED_FULL_RESULT.json sha256 8256cb120376f46bf00b537d2ad665f9d50335e3eed3727a1299129195dfb65b Committed prediction: hierarchical-aif/protocols/B4C01-CORRECTED-FULL-PREDICTION.md @ 28ce738 Maps to: P0/P1 pipeline integrity; constrains the interpretation of P3. No P-level is raised.

1. Executive result

B4C01 ran to completion at the full frozen N = 200 simulations per generator × 5 generators = 1000 simulations, with zero failures and resourceBoundPartial: false. runStatus = ELIGIBLE_FOR_FROZEN_VERDICT.

Verdict NOT_ESTABLISHED. The frozen expectation PASS is REFUTED at full N.

Exactly one criterion failed: failingCriteria = ['M0_EXPONENTIAL:self_win'].

All four parameterised generators recovered their true parameters well inside tolerance. The failure is entirely in self-win, and entirely in the one generator that every other competitor nests.

Per the four-way classification pre-committed in the landing template, this is unambiguously:

CLASS B — self-win / power failure with parameter recovery intact. This is a POWER / ASSAY-RESOLUTION finding, NOT a fitter defect. Reading it as "the B3 pipeline is broken" is a misreading and is forbidden.

2. Results by generator

Parameter recovery is not merely inside tolerance — it is inside it by one to two orders of magnitude. The largest bias is M2's −0.0216 against a 0.1 tolerance (22% of budget); M5's is +0.00063 against 0.15 (0.4% of budget). The fitters are unbiased on their own synthetic data. P1 for the simple-model fitters is in no way implicated.

3. What actually failed, and why it is a statement about the assay

M0_EXPONENTIAL self-won 58 of 200 simulations (0.290) against the frozen > 0.5 threshold.

The mechanism was pre-committed in §7 of the prediction record and the result matches it: M0_EXPONENTIAL is nested inside M1_WEIBULL (k = 1), inside M5_GAMMA (shape = 1), and is a degenerate limit of M3_TWO_TIMESCALE. When the truth is exponential, those three fit it essentially as well; the only price they pay out-of-sample is the estimation variance of one or two extra parameters, of order p / 2n ≈ 0.0006 nats — far below the sampling noise of a motor-equal score over 19 synthetic holdout motors. And M0 must beat all five competitors simultaneously to self-win.

0.290 sits just above the ~0.25 that near-random selection among four indistinguishable models would give. That residual is the true model's real but unresolvable edge.

The assay is not uniformly blind — and that is the sharper finding. The two structurally distinctive generators self-win strongly: M3_TWO_TIMESCALE at 0.935 and M2_LOGNORMAL at 0.885. The design resolves models that differ in shape. It is specifically blind between a model and the competitors that nest it. "Underpowered" understates it; the correct statement is nesting-blind at this holdout size.

What this licenses about the rest of the programme

This is the sanity-floor control that makes other cells interpretable, so its failure mode matters beyond itself. It means: an INCONCLUSIVE/NOT_ESTABLISHED contrast between nested or near-equivalent models in B3 or in the F-side scoring is the expected output of this design, not evidence about the models. It is direct, measured support for the standing caution that underpowered is not equivalence — and it now has a number attached.

It does not license doubting B3's resolved contrasts, nor the adverse M2-over-M3 finding: those involve structurally distinct models, exactly the case this cell shows the assay can resolve.

4. Prediction scorecard — graded against the COMMITTED record, unrevised

The prediction record was committed at 28ce738 while the cell had never been executed at any N.

Six clean hits, one narrow miss, two misses. The misses are recorded, not absorbed:

• Item 7/8 — I underestimated M3. I reasoned from B4C02 that M2 threatens M3 in heavy-tailed worlds, and carried that into M3's own world. Wrong: when the generator is the two-timescale process, M3 self-wins at 0.935 — the highest of any generator. The B4C02 lesson (M2 beats M3 under a three-timescale truth) does not transfer to M3's correctly-specified case, and I over-transferred it.
• Item 8 — the ordering. I placed M5 below M1; observed M1 (0.590) below M5 (0.670). Both are one-parameter shapes competing against each other and against M0; I have no principled account of which should lose more, and the prediction claimed more resolution than my reasoning supported.

The primary structural claim — NOT_ESTABLISHED driven by M0 self-win with recovery intact — was correct, and it was the risky part. The frozen expectation said PASS.

5. Provenance

D2 — the cost lesson, inverted this time

The smoke measurement taken under the conditions the run would actually face beat the cross-cell projection. B4C02's rate was measured while it shared the box differently; B4C01 ran its whole life alongside B4C11. Measure in the regime you will run in — a refinement of D2, and the committed record's promise that "the harness will record its own secondsPerSim so this projection is itself falsifiable" was honoured, and it falsified it.

D4 — the frozen reason text remains wrong, and is superseded not edited

The frozen artifact justified NOT_RUN with "~15–25 min per M4/M7-inclusive competition". This cell fits neither M4 nor M7. corrected_reasons.C01_REASON supersedes it for reporting; the frozen artifact is not edited.

6. Prospectivity — SATISFIED, and the cleanest in the batch

Unlike B4C02 (committed mid-run) and B4C11 (committed at 210/2000), there was no partial evidence of any kind in existence when this prediction was committed. Compare B4C10 and the F-side scoring, both permanently NOT_SATISFIED under D9.

Per D9, the flip PENDING → PROSPECTIVE belongs in the result commit, and the prediction and result must never share a commit. The result is currently uncommitted; that ordering is intact.

7. Lane impact — scoped

D3 closes: the corrected harness ran deterministically to completion with PYTHONHASHSEED unset and 0 failures. D2 remains CLOSING until B4C11 lands.

8. What this cell cannot do

• No mechanism, no biological parity, no active-inference claim. Every dataset is SYNTHETIC; nothing here may be labelled OBSERVED.
• It cannot move P6. It speaks to P0/P1 integrity and constrains the interpretation of P3.
• M4/M7/M8 are skipped by construction, so it says nothing about the mixture, hierarchical or KDE models — including the F-side candidate, which is not entered.
• It does not invalidate B3. The failure is confined to nested near-equivalent models; the distinctive generators self-win at 0.885 and 0.935.
• withinTolerance tests bias on the median, not spread — a real weakness of the frozen criterion, applied as written rather than fixed here.

9. Limitations

Five generators do not span plausible model space. One cohort geometry — every simulated dataset inherits the derived_eligible_1_to_8 structure and its 19-holdout-motor sampling noise. Self-win is a win-rate, not an effect size, and carries no interval; the frozen criterion is a bare threshold comparison, so no CI-bound statement about 0.290 is available and none is made.

NEXT_ACT = supervise PID 26756 (B4C11, 1820/2000, ~1.9 h out; watcher bgd8slocg armed). On landing: sha256sum the result into B4C11_CORRECTED_FULL.sha256, fill reports/templates/B4C11-CORRECTED-FULL-REPORT.TEMPLATE.md into reports/B4C11-CORRECTED-FULL-REPORT.md against committed prediction 897c8ab, route D1, update the P-ladder mapping, run both guards and the suite