B4C02 — Corrected Full-N Prediction Record

PROSPECTIVITY: PENDING_SATISFIABLE (graded by the commit graph, not by prose — D9). Written 2026-07-21 prior to any B4C02 run, and committed 2026-07-21T22:44:31Z in b9b5670, which is before any B4C02 observation exists — the run is still in flight and no result file has been written. If the result lands after that commit, as it must, this cell's prediction is genuinely prospective. This is the only cell in the current batch for which that holds. Gate: H-AIF-G5 · Cell: B4C02_MISSPECIFIED_WORLDS · Priority: HIGHEST (epistemic value)

1. Purpose

B4C02 is the misspecified-world discriminator. It asks whether the B3 adverse result — a simple lognormal (M2) out-predicting the two-timescale mixture (M3) on held-out data — is a generic feature of heavy-tailed dwell shape rather than evidence about any particular mechanism.

It was recorded NOT_RUN with a RESOURCE_BOUND reason claiming 150–250 h. Measured cost is ≈8.7 h (D2). It is the single most decisive piece of missing evidence in the submitted package.

2. Why this cell runs first

Under active-inference action selection, when ambiguity is high the next action should be the one with the highest epistemic value, not the cheapest. B4C10 is cheaper (2.1 h) but resolves less: it completes a replicate count on a cell whose method was never in doubt. B4C02 discriminates between two live explanations of the headline adverse result. It runs first.

3. Design (frozen, unchanged)

Three misspecified generators, each simulating datasets from a world that is not any competitor:

• planned_N = 200 simulations per generator (frozen), 600 total.
• Each simulation: rebuild cohort → fit M0, M1, M2, M3, M5, M6 → score motor-equal NLPD on the simulated holdout → record the winner and whether M2 beats M3.
• M4/M7/M8 are skipped by construction — this is the cell's declared design, and the honest statement of its model set (see D4).
• Frozen criterion: gensWithM2overM3 = #{generators with m2_beats_m3_frac >= 0.5}. >= 2 of 3 → GENERATOR-ROBUST_ADVERSE; otherwise GENERATOR-SPECIFIC.

4. The single deviation from the committed runner

D3 seeding fix only. The committed cell seeds with np.random.default_rng(seed_base + sim + hash(gen_label) % 100000), which is non-deterministic across processes. The corrected harness substitutes seeding.stable_seed(cell_id, base_seed, replicate_index, protocol_version, cohort_id), a SHA-256-derived integer.

Everything else — generators, cohort construction, fitting, scoring, aggregation, criterion — is called through the same b3/b4 functions. No threshold, no N, no criterion, no model set is changed. seed_base = 20260802 is unchanged.

Consequence to state 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 produces different data on every invocation. This run is reproducible; that one was not.

5. Hypotheses held alive

6. Pre-committed outcome mapping

Falsifier of the frozen prediction: gensWithM2overM3 <= 1 at full frozen N.

7. What this cell cannot do, whatever it returns

• 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 interpretation of P3.
• A GENERATOR-ROBUST_ADVERSE result is not a defeat of the motor-stack AIF model, which is not entered in this competition. It is a statement about what the M2/M3 contrast can support.
• Because M4/M7/M8 are skipped, this cell says nothing about the mixture, hierarchical, or KDE models.

8. Expected runtime

≈8.7 h (600 sims × 52.1 s measured). Frozen N; no reduced protocol required.

9. Wording

Allowed: "B4C02 corrected full run supports / weakens / does not establish the shape-artifact explanation of the adverse M2-over-M3 result within its frozen scope (3 generators, 6 simple competitors, motor-equal NLPD, single 19-motor cohort)."

Forbidden: "mechanism demonstrated" · "biological parity" · "active inference proved" · "full flagellum parity" · "M2 is the UNI model" · "M3 is vindicated" · any statement that this cell tests the motor-stack AIF model.