B4C01 — Corrected Full-N Prediction Record

PROSPECTIVITY: decided by the commit graph, never by this sentence (D9). This record does not assert its own standing; the ordering is asserted by test_b4c01_prediction_was_committed_before_any_observation and is recomputed from git on every test run.

What is directly verifiable at the time of writing, and is all that is claimed here: B4C01 has never been executed at any N. The frozen artifact records status: NOT_RUN, actual_N: 0; there is no smoke test, no partial run, and no result file for this cell anywhere in the repository. If the commit graph then places this record ahead of the first B4C01 observation, this cell will have the cleanest standing in the batch — B4C02's record was committed mid-run and B4C11's at 210/2000 replicates, so both had partial evidence in existence, whereas here there is none of any kind to have been influenced by.

Gate: H-AIF-G5 · Cell: B4C01_SYNTHETIC_PARAMETER_RECOVERY · Lane: B Defects routed: D3_HASH_SEED_NONDETERMINISM (corrected here) · D2 (cost re-measured) · D4 (the frozen reason text is wrong about this cell's model set)

1. Purpose

B4C01 is the correct-specification sanity floor. It asks: when data are generated from a competitor with its own frozen parameters, does the B3 pipeline (a) recover those parameters within tolerance, and (b) let that model win its own competition more often than not?

It is the control that makes every other B4 cell interpretable. If the pipeline cannot recover a model from its own synthetic data, no adverse or favourable result elsewhere can be trusted.

It has never been run. The frozen artifact records NOT_RUN with a RESOURCE_BOUND reason.

2. Planned run

3. The one and only deviation: the D3 seeding fix

The committed cell_C01 seeds each simulation with:

rng = np.random.default_rng(seed_base + sim + hash(gen) % 100000)

hash(str) is randomized per process when PYTHONHASHSEED is unset, so the committed cell produces different synthetic data on every invocation and can never satisfy a byte-determinism gate. This is defect D3, identical in kind to the one corrected for B4C02.

The corrected harness substitutes seeding.stable_seed(cell_id, base_seed, replicate_index, protocol_version, cohort_id), a SHA-256-derived integer. Nothing else changes. Every science function is still called through the frozen b3/b4 modules: b4._simulate_from_model, b4._build_cohort_from_events, b3.fit_simple_models, b3.fit_m6, b3.scoring_params, b3.nlpd_per_event, b3.aggregate_motor_equal.

Consequence, stated plainly: because the seed derivation changes, this run is not bit-comparable to a hypothetical run of the committed cell. It could not have been anyway — the committed cell is non-deterministic by construction. This run is reproducible; that one was not.

4. Frozen criteria — unchanged, not restated loosely

True parameters are the frozen B3 fits on derived_eligible_1_to_8, read from audits/phase-b/b3-model-competition-result.json (sha256 5d7a0589…):

Verdict rule (frozen, verbatim from cell_C01):

PASS iff for every generator: withinTolerance AND self_win_frac > 0.5. For M0_EXPONENTIAL (no parameters) only self_win_frac > 0.5 applies. Otherwise NOT_ESTABLISHED.

• Scoring: motor-equal NLPD on each simulation's own synthetic holdout.
• Competitors scored per simulation: M0, M1, M2, M3, M5, M6_SEMI_MARKOV_STATE_DEPENDENT.
• M4_MIXTURE_K3, M7_HIERARCHICAL_MOTOR, M8_EMPIRICAL_KDE are SKIPPED by construction. This is the cell's declared design and is recorded in its own output as a runtime concession.
• withinTolerance is judged on the median recovered parameter across simulations, i.e. it tests bias, not spread. A model could be wildly variable per simulation and still pass this criterion. That is a real weakness of the frozen criterion and it is not being fixed here — the frozen rule is applied as written, and the spread is reported alongside it.

D4 — the frozen reason text is wrong about this cell

The frozen artifact justifies NOT_RUN with "~1000 refits × ~15–25 min per M4/M7-inclusive competition = ~250–400 h". This cell does not fit M4 or M7. The corrected reason (corrected_reasons.C01_REASON) supersedes it for reporting; the frozen artifact is not edited.

5. Expected runtime — MEASURED, not asserted (D2)

The completed B4C02 run used the same per-simulation fit path (fit_simple_models + fit_m6 + scoring of the same 6 competitors):

B4C02 measured: 29 409.096617221832 s over 600 simulations = 49.015 s/sim
B4C01 projection: 1000 simulations x 49.015 s = 49 015 s = 13.62 h

This supersedes both the frozen 250–400 h claim (overstated ~18–29×) and the earlier 14.5 h estimate derived from component timings. The projection is anchored to a completed full-N run of a structurally identical cell, not to a component estimate. The harness will record its own secondsPerSim so this projection is itself falsifiable.

6. Pre-committed outcome branches

7. Directional prediction — committed, and I expect to be wrong about the headline

The frozen expectation is PASS. I predict NOT_ESTABLISHED (branch b).

Parameter recovery: I predict ALL FOUR parameterised generators pass withinTolerance. Maximum-likelihood estimates are consistent, the tolerances are generous relative to the parameter scales (e.g. 0.1 on k = 0.625 is 16%), and the criterion is on the median over 200 simulations, which suppresses per-simulation variance almost entirely.

Self-win: I predict M0_EXPONENTIAL is the failure point, with self_win_frac in [0.20, 0.50], most likely 0.30–0.40.

Mechanism for that prediction. 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 competitors fit it essentially as well; the only penalty they pay out-of-sample is the estimation variance of one or two extra parameters, of order p / 2n ≈ 0.0006 nats. That signal is far smaller than the sampling noise of a motor-equal score over 19 synthetic holdout motors. So the argmin over {M0, M1, M5, M3} is close to a coin toss among four near-indistinguishable models, and M0 must beat all five competitors simultaneously to self-win. Near-random selection among ~4 equivalent models puts self_win_frac near 0.25–0.40, below the 0.5 threshold.

Secondary risk: M5_GAMMA, self_win_frac in [0.40, 0.70]. Gamma with shape = 0.512 and Weibull with k = 0.625 produce very similar densities on this support; M1 is its closest rival.

Predicted ordering of self_win_frac, lowest to highest: M0_EXPONENTIAL < M5_GAMMA < M3_TWO_TIMESCALE < M1_WEIBULL < M2_LOGNORMAL. I predict M2_LOGNORMAL > 0.80 — sigma = 1.578 is a distinctive heavy tail no competitor here mimics well. I predict M3_TWO_TIMESCALE in [0.50, 0.80], with M2 its main threat: B4C02 showed M2 beating M3 in 94% of simulations from a three-timescale world, and a two-timescale world is the adjacent case.

What would make me wrong in the interesting direction. If M0's self_win_frac comes in well above 0.5, the out-of-sample penalty for unnecessary parameters is materially larger than p / 2n at this sample size — which would say something useful about how sharply this design punishes over-parameterisation, and would strengthen every B3 contrast.

8. Falsifier

• The frozen prediction (PASS) is falsified if any generator fails either condition.
• My prediction is falsified item by item: a PASS verdict; or M0_EXPONENTIAL with self_win_frac > 0.5; or any parameterised generator failing withinTolerance; or the predicted ordering of self_win_frac coming out different. Each is a recorded miss.

9. What this cell cannot do, whatever it returns

• It cannot establish biological parity, mechanism, or active inference. Every dataset here is synthetic; nothing in this cell may be labelled OBSERVED.
• It cannot promote any model to "the UNI model." M2_LOGNORMAL remains an adversarial baseline.
• A PASS does not vindicate M3 and does not support the F-side motor-stack model, which is not entered in this competition.
• Because M4/M7/M8 are skipped, this cell says nothing about the mixture, hierarchical, or KDE models.
• It cannot move P6. It speaks to P0/P1 pipeline integrity and constrains the interpretation of P3.
• A NOT_ESTABLISHED driven by nested-model near-equivalence is a power statement about a 19-motor holdout, not a defect finding. Reporting it as "the B3 pipeline is broken" would be a misreading, and is forbidden.

10. Wording

Allowed: "B4C01 at full frozen N supports / does not support parameter recovery and self-win for the simple model set on synthetic data, within its frozen scope (5 generators, 6 competitors, motor-equal NLPD, single 19-motor cohort geometry)." · "sanity floor met / not met" · "not established" · "power-limited at this holdout size".

Forbidden as claims (canonical list in claim_guard.FORBIDDEN and the H-AIF contract; referenced, not re-transcribed, so a wrapped catalogue cannot read as an assertion): any parity, mechanism, active-inference, or "solved" claim; "M2 is the UNI model"; "M3 is vindicated"; any statement that this cell tests the motor-stack AIF model or the mark process.

11. Required pre-run checks

• Corrected harness reproduces the frozen generator list, order, and seed_base = 20260801.
• stable_seed substitutes only for hash(gen) % 100000; the + sim term and every other arithmetic input are unchanged.
• True parameters are read from the frozen B3 artifact at run time, never hard-coded.
• Tolerances, the > 0.5 self-win threshold, and the PASS/NOT_ESTABLISHED rule are unmodified.
• M4/M7/M8 remain skipped and the omission is recorded in the result.
• NO FLOOR — a non-finite log density increments failed; it is never replaced or clipped.
• Frozen artifacts under audits/** are untouched; the harness only reads them.
• Deterministic JSON output (indent=1, sort_keys=True, LF) with a printed sha256.

12. NEXT_ACT

NEXT_ACT = commit THIS record first (no B4C01 observation exists yet, so its prospective standing
           is clean); then build hierarchical-aif/scripts/run_c01_corrected_full.py, smoke-test it
           to a scratch path, and launch at the frozen N_sim = 200 x 5 generators