F-Side Motor-Stack — Held-Out Scoring Result

Gate: H-AIF-G7 · Lane: D (motor-stack AIF) · Date: 2026-07-22 Prediction record (WRITTEN before this run; NOT committed before it — prospectivity NOT_SATISFIED, see §2 and D9): hierarchical-aif/protocols/F-SIDE-MOTOR-STACK-SCORING-PREDICTION.md Result: hierarchical-aif/results/motor_stack_aif/F_SIDE_MOTOR_STACK_SCORING_RESULT.json sha256 b3b12720f32c0aee3bfa456f52ae0901976e59e3b43c0f2690fa7a17386ab297 (22 789 bytes) Maps to existing ladder: P3 held-out predictive, duration-only scope. No P-level is raised.

1. Executive result

The constrained F-side motor-stack model was fitted on the 80 training motors and scored on the 19 frozen holdout motors under the frozen motor-equal NLPD rule. Outcome: pre-committed branch (B), NOT_ESTABLISHED, against the control and against every serious adversary. Two contrasts resolved above zero, both against weak adversaries (M0_EXPONENTIAL, M5_GAMMA).

The single most important number in this report is not a contrast. It is this:

The F-side model reproduces the incumbent hierarchical model to within 2.5e-7 nats. At the resolution this study can achieve it is not a new model; it is a re-derivation of M7. That is the honest headline, and it is an adverse result for the hypothesis that the constrained motor stack buys predictive power. It is also a strong implementation receipt: two independently written hierarchies agree to 7 significant figures.

2. Provenance and integrity

Independent-oracle consistency check — PASS, residual exactly 0.0

B3 stores only aggregated scores; there are no per-motor arrays in the frozen record, so every per-motor array here had to be recomputed. That is precisely why this check is mandatory and not optional. Three frozen models were re-derived from their recorded parameters and compared against their published motor-equal values:

Tolerance 1e-12, and the observed residual is exact zero. The harness halts rather than warning if this check fails; it never proceeds to a verdict on an unvalidated scale convention. This is what licenses the seconds-vs-normalised handling for every other model in the table.

3. Held-out leaderboard (motor-equal NLPD, seconds scale, lower is better)

Placed into the full frozen 9-model B3 leaderboard, the F-side model would sit 5th of 9, between M6_SEMI_MARKOV_STATE_DEPENDENT (3.4298889710) and M1_WEIBULL (3.4333068484), and below M2_LOGNORMAL, M8_EMPIRICAL_KDE, M4_MIXTURE_K3 (3.4241154309) and M6.

A point-estimate ranking is not a verdict. The CI-bound verdicts follow.

4. Contrasts — CI-bound verdicts

Convention: contrast = S(reference) - S(F_MOTOR_STACK). Interval entirely above 0 ⇒ the F-side model is better; entirely below 0 ⇒ the reference is better; contains 0 ⇒ NOT_ESTABLISHED.

Branch determination, per the pre-committed table:

• Not branch (A). Branch A required the interval against the control and every adversary to lie entirely above 0. Four of six contain 0.
• Branch (B) is the governing outcome for the control and for all three serious adversaries (M2, M8, M1). NOT_ESTABLISHED — this is not equivalence and not "no difference."
• Not branch (C). No adversary's interval lies entirely below 0, so H_FSIDE_DURATION is not falsified either. M2_LOGNORMAL and M8_EMPIRICAL_KDE lead on point estimate but their intervals cross 0.
• Two contrasts resolved above 0 against the two weakest adversaries. Beating a unit exponential and a gamma is a floor check, not a scientific claim.

The M1_WEIBULL contrast is the informative one

M1 is the tau -> 0 nesting control: it is exactly what the F-side model collapses to if the population spread does no work. The contrast is +0.000615 nats with a half-width of 0.0111 — the tightest contrast in the table, and centred essentially on zero. The entire hierarchical apparatus — a population prior over log-shape with per-motor latents integrated by 33-node Gauss-Hermite quadrature — buys approximately nothing over a single pooled Weibull on held-out duration prediction at this sample size. The fitted tau = 0.1837 is small, consistent with that reading.

5. Prediction scorecard — the pre-committed §8 predictions, scored

The prediction record committed eight specific claims before the numbers existed. Scored honestly:

Seven of eight, with one recorded miss. The miss is item 6 and it is methodologically instructive, so it is not buried:

The M5_GAMMA contrast resolved with a half-width of 0.0250, well below the ~0.042 nat resolution floor the protocol used as its rule of thumb. The floor is a heuristic derived from the narrowest frozen B3 contrast, not a bound. A contrast can resolve below it when the per-motor difference is unusually consistent in sign across motors, because the paired motor-cluster bootstrap is sensitive to consistency, not just magnitude. The protocol's §10 anticipated this in words ("unless the per-motor difference is unusually consistent") but §8 still predicted the wrong verdict. Recorded as a miss.

Calibration caveat, stated so the scorecard is not over-read. The prediction was formed with the frozen published holdout scores of the competitors already in hand (legitimately — that channel was spent by B3). Given a predicted F-side point location, most contrasts follow by arithmetic. The genuinely risky commitment was item 1, the point location itself, and the reason it landed to 2.5e-7 is the finding of §1: the F-side model and M7 are the same model to this resolution. The scorecard measures the reasoning, not new evidence.

6. Adverse results — retained, in the headline

1. M2_LOGNORMAL out-predicts the F-side motor stack on point estimate (3.4093 vs 3.4327, contrast -0.023378). The interval crosses 0, so this is NOT_ESTABLISHED, not a refutation — and equally not a defence. The standing B3 adverse finding (M2 over M3 by ~0.0369 nats event-pooled) is extended, not overturned: a two-parameter lognormal remains the best point performer on this cohort, ahead of every mechanistic candidate including this one.
2. M8_EMPIRICAL_KDE, a purely empirical model with no mechanism at all, also out-predicts the F-side motor stack on point estimate (3.4225 vs 3.4327).
3. The hierarchy buys nothing measurable over M1_WEIBULL (§4).
4. The F-side model is not distinguishable from the incumbent M7 (§1).
5. M2_LOGNORMAL is an ADVERSARIAL BASELINE, never the UNI model. Its performance is retained as an adverse result about the mechanistic candidates, not promoted to a claim about lognormal mechanism.

7. What this establishes, and what it does not

Establishes (scoped): on the derived_eligible_1_to_8 cohort of this single passive dataset, at 19 holdout motors, under motor-equal NLPD on the seconds scale, the constrained F-side motor stack predicts held-out durations better than M0_EXPONENTIAL and M5_GAMMA by a CI-bound margin, and is NOT_ESTABLISHED against the current design and against every serious adversary. The implementation is deterministic, censoring-correct, no-floor, motor-equal, and reproduces three frozen models to exact zero residual.

Does not establish — none of this moves:

• No mechanism. Predictive ordering is never promoted to mechanism.
• No biological parity at any level. Full biological parity is not a current status.
• No active-inference claim. The dataset is passive and the action set is empty.
• No G-side policy claim. expected_free_energy does not exist in this package, a test enforces its absence, and it must not be added. G-side remains design-only until intervention or transfer.
• Nothing about the mark process. nextStateN/direction/jump were not read (D5).
• Nothing about B4C11, whose U4_OK remains withdrawn (D1). The corrected C11 run is in flight; this result does not speak to it.
• P4 transfer, P5 intervention, P6 structural/mechanistic, P8 full verdict are untouched in either direction. P8 remains FULL_PARITY = false; the first unsatisfied level is still P4 transfer.

8. Limitations

• n = 19 holdout motors. The 233 holdout events are not 19-fold more information; events within a motor are not independent replicates. Most contrasts are expected to be inconclusive, and they were.
• Underpowered is not equivalence. Every NOT_ESTABLISHED above is reported with its interval and half-width. Replicates were not increased after seeing a width, and must not be.
• One cohort, one study, one species context (E. coli behavioural evidence). No transfer.
• Right-censored events are excluded by the frozen cohort, so the F-side censoring branch is correct-by-test but unexercised by this scoring run. That is a real gap in coverage of the implemented likelihood, not a property of the result.
• The intervals are percentile, not BCa. The frozen B3 verdicts used BCa (intervalUsed). The half-widths quoted here belong to the percentile intervals and are labelled as such (D7).
• M4_MIXTURE_K3, M6_SEMI_MARKOV_STATE_DEPENDENT and M7_HIERARCHICAL_MOTOR were not contrasted here — only their published aggregates are available, and a paired motor-cluster bootstrap needs per-motor arrays. Their published point scores are quoted for placement only. This was a genuine coverage gap. It is now CLOSED - see reports/M4-M6-M7-PER-MOTOR-CONTRASTS-REPORT.md (post-hoc extension, 751a59ef...). All three per-motor arrays were re-derived from their frozen fits and reproduce their published aggregates at exact zero residual. Outcome: neither M4 nor M6 resolves against the candidate (NOT_ESTABLISHED both, nominal and Bonferroni), and the M7 contrast exposed D10 - the frozen CI rule has no minimum-effect-size guard and "resolved" a 2.5e-07-nat difference. The candidate ranks 5th of 10 combined.
• A RuntimeWarning: overflow encountered in power is emitted from weibull_log_survival during the fit. It is benign and by design: extreme quadrature nodes produce a non-finite density, which _check converts to NonFiniteLogDensity, which the marginal maps to -inf (zero mass for that node). No floor is applied. It is recorded here rather than silenced.

9. Gate status

P3 duration-only: unchanged. B3 stands. The F-side candidate joins the leaderboard without displacing anything and without being displaced by a CI-bound margin.

10. What would make LANE D true

Framing the next receipt as "what would make this lane true", not "why it fails":

1. Per-motor arrays for M4, M6, M7 so the three models that out-score or match the F-side candidate can be contrasted under the same paired bootstrap. This is the cheapest highest-value next step and needs no new data.
2. More independent motors. The binding constraint is 19 holdout motors, not model quality. A second dataset with independent motors would move P4 and sharpen every contrast here.
3. A model that predicts something M1_WEIBULL cannot. The hierarchy currently buys nothing; an F-side model earns its complexity only by beating its own tau -> 0 limit.
4. Censored observations, to exercise the implemented censoring branch on real data.
5. Intervention data for any G-side claim. No amount of passive data substitutes; the action set is empty, which is structural.

NEXT_ACT = monitor B4C02 (running) and B4C11 (running, launched 2026-07-22T01:40:03Z); on each completion sha256 the result, write its report against its pre-committed prediction record, and update the defect closure ledger (D1 for C11, D2 for C02)