M4 / M6 / M7 Per-Motor Contrasts — closing the F-side coverage gap

Gate: H-AIF-G7 (extension) · Lane: D · Date: 2026-07-22 Result: hierarchical-aif/results/motor_stack_aif/M4_M6_M7_PER_MOTOR_CONTRASTS_RESULT.json sha256 751a59ef45c8aecd1bdbe5fb5ef645423572a24d1691fb199d1b3b33fb8d4dbb Status: POST_HOC_COVERAGE_EXTENSION — see §2. No P-level moves on this artifact.

1. What this closes, and what it found

The F-side scoring report recorded a coverage gap in its own §8:

M4_MIXTURE_K3, M6_SEMI_MARKOV_STATE_DEPENDENT and M7_HIERARCHICAL_MOTOR were not contrasted — only their published aggregates are available, and a paired motor-cluster bootstrap needs per-motor arrays. This is a genuine coverage gap: M4 and M6 both out-score the F-side model on point estimate and neither has been contrasted against it under a CI-bound test.

That gap is now closed. B3 stores only aggregates, so all three per-motor arrays were re-derived from their frozen fitted parameters through the frozen scoring functions — no refitting, no new data, no held-out mark channel.

Headline: neither M4 nor M6 resolves against the candidate. And the M7 contrast exposed a defect in the verdict rule itself (D10) that is more important than any of the three contrasts.

2. Post-hoc status — declared, not hidden

The pre-registered F-side protocol named its competitor set in advance: CONTROL_CURRENT (M3) plus M0/M1/M2/M5/M8. M4, M6 and M7 were not in it. The decision to add them was taken after seeing the candidate rank 3rd of 7 there.

A Bonferroni companion at alpha/9 is reported as a sensitivity; the frozen convention decides on the nominal 95% interval. These verdicts are POST_HOC_EXPLORATORY and do not carry the standing of the six pre-registered contrasts.

3. Oracle gate — PASS, residual exactly 0.0 on all three

B3 stores no per-motor arrays, so an unvalidated recomputation would be worthless. The gate HALTS the script and emits no verdict on failure.

Two reconstructions were non-obvious and are therefore worth naming, because the gate is what proves they were right: M4's scoring dict had to be rebuilt from canonical ({rates, weights} — the result JSON carries no m4params key), and M6's per-state parameters arrive from JSON with string keys and must be re-keyed to int because the frozen scorer indexes by integer state. Either mistake would have produced plausible-looking numbers; both were caught by the exact-zero requirement.

Six frozen models have now been reproduced to exact zero residual (M3, M0, M8 in the F-side run; M4, M6, M7 here). That is a strong P1 implementation receipt for the scoring path.

4. Contrasts

Convention: contrast = S(reference) - S(F_MOTOR_STACK); above 0 ⇒ candidate better.

M4 and M6 lead on point estimate but neither resolves. The gap the F-side report flagged is closed with a null: at 19 holdout motors these differences are not resolvable. NOT_ESTABLISHED is not equivalence.

M6's entire interval lies inside ±0.042 nats. No scientifically material difference is detectable in either direction. That is a stronger statement than a bare NOT_ESTABLISHED — but it is not a formal equivalence result, because the floor is a heuristic derived from the narrowest frozen B3 contrast and was never pre-specified as an equivalence margin. A test enforces that no equivalence language attaches to it.

Combined leaderboard: the candidate ranks 5th of 10.

5. D10 — the verdict rule has no minimum effect size

This is the real finding, and it is adverse to the methodology, not to any model.

The M7 contrast was resolved by the frozen rule: the interval [+1.60e-07, +3.37e-07] genuinely excludes 0. But the effect is 2.507e-07 nats — about 168 000× below the 0.042-nat resolution floor.

Root cause, verified from first principles by test: a paired motor-cluster bootstrap resamples motors, so it resolves a difference of any magnitude provided the sign is consistent across them. Here 16 of 19 per-motor differences were positive with a mean of +2.5e-07, because the F-side hierarchy is M7 to numerical precision — the same model reached by a different quadrature (33-node Gauss-Hermite vs M7's own), differing only in float noise.

Reporting that as "the F-side candidate beats M7" would be truth laundering: a numerically identical model presented as a winner on the strength of consistent float noise.

Handling. The frozen verdict is reported verbatim and unaltered — it is not softened, suppressed, or re-thresholded. Every contrast now additionally carries a scientificReading (MATERIAL / SUB_FLOOR_EFFECT / SCIENTIFICALLY_NULL) and a reportableAsAWin flag which is false whenever a resolved verdict sits below the floor. This changes no frozen threshold, criterion, or interval.

This also explains the recorded miss in the F-side prediction scorecard. Item 6 predicted M5_GAMMA would be NOT_ESTABLISHED because its +0.031 effect sits below the ~0.042 floor; it resolved instead. Same mechanism, milder form: the floor is about consistency, not magnitude. D10 supersedes the "resolution floor" as a predictor of what will resolve — the floor tells you what is scientifically material, not what the bootstrap will call.

6. What this does not establish

• No mechanism, no biological parity, no active-inference claim. Duration-only, one cohort, 19 holdout motors.
• No P-level moves. These contrasts are post-hoc and exploratory.
• The candidate did not beat anything here. Two nulls and one artifact.
• M7 is not displaced, and the F-side model is not shown superior to it — the two are the same model at this resolution, which was already the headline of the F-side report.
• The retained adverse finding stands: M2_LOGNORMAL (3.4093) and M8_EMPIRICAL_KDE (3.4225) still lead the candidate (3.4327) on point estimate.

7. Limitations

• Post-hoc selection, family size 9, Bonferroni companion reported (no contrast's verdict changed under it).
• Percentile intervals, not BCa. The frozen B3 verdicts used BCa (intervalUsed); every half-width here is labelled percentile (D7).
• 19 holdout motors. The binding constraint throughout.
• The SCIENTIFICALLY_NULL classification uses a heuristic floor, not a pre-specified equivalence margin.
• M7's parameters (k = 0.6596322379287862, tau = 0.18372185667134974) versus the F-side fit (exp(mu) = 0.6596326697…, tau = 0.18372082607…) differ in the 6th–7th significant figure. The contrast is measuring that, and nothing else.

NEXT_ACT = python -m pytest hierarchical-aif/tests/motor_stack_aif -q && watch PID 26756 (B4C11) and PID 32988 (B4C01) to completion