F-Side Motor-Stack Held-Out Scoring — Prediction Record

PROSPECTIVITY: NOT_SATISFIED — not retroactively repairable (D9). This file is UNTRACKED, and the scoring result already exists (2026-07-22T01:47:14Z), so committing it now would post-date the observation. What is true and is all that is claimed: it was written prior to any held-out scoring run of the F-side motor-stack model. The harness that will execute it, hierarchical-aif/src/motor_stack_aif/compare.py, exists and has been exercised TRAIN-ONLY; compare.run_comparison() has not been called.

Gate: H-AIF-G6 (F-side observational projection) · Lane: D (motor-stack AIF) Parity ladder: P3 duration-only only. P5/P6/P8 untouched in either direction. Status on commit: PENDING. It becomes PROSPECTIVE only in the result commit, and only if this prediction commit is a proven strict ancestor of the result's introduction.

1. Claim under test

H_FSIDE_DURATION. The constrained F-side motor stack — a two-parameter population prior over per-motor log-shape, with the per-motor latent integrated out — predicts held-out dwell durations better than CONTROL_CURRENT (M3_TWO_TIMESCALE, the frozen B3 reference model) and better than every simple adversary, by a CI-bound margin on the frozen split.

This is a claim about held-out duration density only. It is not a claim about mechanism, about the mark process, about active inference, or about any biological parity level.

2. Experimental unit, split boundary, sample size

Why held-out access is permitted here. The held-out duration channel was already prospectively spent by B3, and duration-only held-out scoring is the explicitly allowed reuse. The held-out mark channel was burned on 2026-07-21 (D5) and is not touched by this protocol; that quarantine is unaffected by this run in either direction.

Right-censoring limitation, stated up front. The frozen cohort contains no right-censored events, so the censoring branch of the F-side likelihood is unexercised on this cohort. Any result here says nothing about censored-dwell behaviour.

3. The model and its equations

Lmotor-5   population prior over log-shape     eta_m ~ Normal(mu, tau^2)          2 free params
Lmotor-4   per-motor latent log-shape          k_m = exp(eta_m)                   INTEGRATED OUT
Lmotor-3   occupancy state N_i                 y_i = durationS_i / scale_N[N_i]   FROZEN scale_N
Lmotor-2   kinetic mode                        NOT INSTANTIATED
Lmotor-1   hazard / survival                   mean-one Weibull(k_m)

Per-event held-out predictive density (the primary scored quantity):

p(y_i) = INT  f_Weibull(y_i ; k = exp(eta), mean-one)  ·  Normal(eta ; mu, tau^2)  d eta

evaluated by 33-node Gauss-Hermite quadrature. Deterministic; no sampling anywhere.

Total free parameters: 2. (mu, tau). Nothing else is estimated.

Why the per-motor latents are integrated, not estimated. There are 793 training events across 80 training motors — a median of roughly 7 events per motor. Estimating 80 free per-motor shapes would fit 80 numbers to ~7 observations each, and the corrected resolution floor on this cohort is about 0.042 nats (motor-equal contrast half-width). Capacity at that granularity is not identifiable and could not be scored honestly. Integrating the latent keeps the parameter count independent of the motor count, which is the whole point of the constraint. Adding Lmotor-2 policies would add unidentifiable capacity, not testability.

4. Competitor set — why each is a serious adversary

CONTROL_CURRENT is M3_TWO_TIMESCALE, the frozen B3 reference model. It is not refitted. Its parameters [0.3933559993214189, 0.44485933051063775] are read verbatim from audits/phase-b/b3-model-competition-result.json and scored through the frozen B3 scoring functions.

M4_MIXTURE_K3, M6_SEMI_MARKOV_STATE_DEPENDENT and M7_HIERARCHICAL_MOTOR are not re-scored here (their fits are not pinned in a form this harness can reconstruct without refitting). Their frozen published numbers remain on the record and are not superseded by anything in this run.

5. Scoring rule, aggregation, contrast convention, CI

Interval-type honesty (D7). Every interval reported by this harness is a percentile interval and every reported width or half-width belongs to that percentile interval. The frozen B3 artifact's width field was the percentile companion while its verdicts used BCa; this harness does not repeat that ambiguity.

Two F-side scores are reported, and only one is contrasted.

• MARGINAL_PER_EVENT (PRIMARY, the one contrasted): each holdout event scored alone under the latent-integrated marginal — the same information set every competitor receives.
• JOINT_PER_MOTOR / n (SECONDARY, never contrasted): the motor's events share one latent and therefore inform each other. It is a different scoring rule and is reported only so the difference is visible rather than buried.

6. Mandatory independent-oracle consistency check (HALT condition)

B3 stores only aggregated scores — there are no per-motor arrays in the frozen record — so every per-motor array used here is recomputed. The check is therefore mandatory, not optional.

Before any verdict, the harness must reproduce B3's published motor-equal NLPD for the three parameter-pinned models, so that any residual is a convention error and nothing else:

Tolerance: 1e-12 absolute. Justification: the oracle reads parameters verbatim from the frozen JSON (Python float repr round-trips exactly), rebuilds the cohort with the frozen Cohort class in the frozen event order, and takes the normalised-space log density from the frozen holdout_lognorm. Every floating-point operation and every summation order therefore matches the recorded run, so the expected residual is exactly 0.0; 1e-12 on a value of ~3.43 is ~3e-13 relative, about 1000 ULP of pure last-place headroom.

This tolerance will not be loosened to make the check pass. If the residual exceeds it, the harness HALTS, no verdict is emitted, and the discrepancy is the reported finding.

Units are the thing most likely to go silently wrong. B3's published NLPD is on the SECONDS scale (it includes +log(scale_N) per event); score.score_motor_stack returns NLPD on the NORMALISED-y scale. That constant cancels in a contrast and does not cancel in an absolute score, so the failure mode is every contrast looking right while every absolute number is wrong by ~2 nats. The scale conversion is a single named function, and a test mutates it and asserts the oracle check FAILS — a check that cannot fail is worthless.

7. Pre-committed outcome branches

Branches are evaluated per contrast; a run can be A against M0_EXPONENTIAL and C against M2_LOGNORMAL at the same time, and both are reported.

8. Directional prediction — committed, and I am willing to be wrong

Evidence used to form it, all of it legitimately available now: the frozen published holdout numbers for the competitors (already spent, already on the record), and the training-only in-sample behaviour of the F-side fit. The F-side held-out score has not been computed.

Training-only inputs to the prediction (in-sample, 80 train motors, seconds scale, from the train-only smoke run): F-side 3.4081198592824875, M1 3.4008238291287873, M2 3.339864483642674, M5 3.473193289623917, M0 3.7103720252098027. Fitted F-side mu = -0.41607215987582913, tau = 0.18372082607308418, train NLL 575.6701064153622 (vs M3 537.0211284007612, M2 509.28782032664856, M1 583.5856062645818).

I predict outcome branch (B/C), not (A). Specifically:

1. Point location. F-side holdout motor-equal NLPD will fall in [3.425, 3.442], and within ±0.005 of the frozen M7_HIERARCHICAL_MOTOR value 3.4326925889658892 — because both integrate a per-motor Weibull shape and differ mainly in parameterisation and node count.
2. vs M0_EXPONENTIAL: contrast ≈ +0.115 nats → RESOLVED_ABOVE. This is the only contrast I expect to resolve.
3. vs M3_TWO_TIMESCALE (control): |contrast| < 0.010 → NOT_ESTABLISHED.
4. vs M1_WEIBULL (the tau -> 0 nesting control): |contrast| < 0.005 → NOT_ESTABLISHED. The hierarchy will buy almost nothing. tau = 0.1837 is small.
5. vs M2_LOGNORMAL: contrast ≈ -0.024 (M2 better) → NOT_ESTABLISHED, because 0.024 sits below the ~0.042 floor. RESOLVED_BELOW is a live alternative if the per-motor difference is unusually consistent, and would be branch C.
6. vs M5_GAMMA: contrast ≈ +0.031 → NOT_ESTABLISHED (below the floor).
7. vs M8_EMPIRICAL_KDE: contrast ≈ -0.010 → NOT_ESTABLISHED.
8. Strongest adversary will be M2_LOGNORMAL (frozen holdout motor-equal 3.4093141565673215, lowest of the adversary set).

Reasoning. In-sample on the training motors the F-side model is already slightly worse than plain M1_WEIBULL on motor-equal aggregation (3.4081 vs 3.4008) and clearly worse than M2 (3.3399), despite a better pooled training NLL than M1. A model that does not win in-sample under the primary aggregation will not win out-of-sample under it. The fitted tau = 0.184 is small, so the population spread is doing little work. The honest expectation is a model that is competitive with, not superior to, the incumbent — and that the standing lognormal adverse result survives.

What would make me wrong in the interesting direction: if between-motor shape heterogeneity is much stronger among the 19 holdout motors than among the 80 training motors, the integrated latent would pay off out-of-sample in a way it cannot in-sample, and contrast 4 or 5 could resolve above 0.

9. Falsifier

H_FSIDE_DURATION is falsified if any adversary's contrast interval lies entirely below 0 — i.e. a simpler or purely empirical model out-predicts the motor stack on held-out durations by a CI-bound margin. It is not established (not falsified, not supported) if any required interval contains 0.

The directional prediction of §8 is falsified item by item: any listed verdict that comes out different, or a point location outside [3.425, 3.442], is a recorded miss.

10. Power — stated honestly, before the numbers

• n = 19 holdout motors. That is the sample size for every contrast. The 233 events are not 19-fold more information: events within a motor are not independent replicates.
• The corrected motor-equal resolution floor on this cohort is a half-width of about 0.042 nats (the narrowest frozen B3 motor-equal contrast is M4_MIXTURE_K3 at 0.083461; the frozen width field is the percentile companion, not the BCa width — D7).
• Detectable: motor-equal separations of roughly 0.05 nats and larger, and smaller ones only if the per-motor difference is unusually consistent (a paired design with low between-motor variance can resolve a small mean shift).
• Not detectable: the ~0.02–0.03-nat gaps that separate the leading models in the frozen record. Most contrasts in this run are expected to be inconclusive.
• Underpowered is not equivalence. A zero-crossing interval will be reported as NOT_ESTABLISHED with its half-width, never as "no difference", "equivalent", or "on par".
• Replicates will not be increased after seeing a width. N_BOOT = 2000 and n = 19 are fixed here, before the run. The bootstrap replicate count is a numerical-precision knob and does not change the 19-motor information content.

11. What this CANNOT establish, even if branch (A) fires

• No mechanism. Predictive superiority is never promoted to mechanism. A better density is a better density.
• No biological parity at any level. FULL_PARITY remains false; this run touches P3 duration-only and nothing else.
• No active-inference claim. The dataset is passive and the action set is empty. That is structural, not a sample-size limitation.
• No G-side policy claim. expected_free_energy does not exist in the package and a test enforces its absence. It must not be added.
• Nothing about the mark process. P5/P6/P8 are untouched. The D5/D6 quarantines are unchanged in both directions.
• Nothing about B4C11, whose submitted U4_OK remains withdrawn under D1.
• Nothing about censored dwells (excluded from this cohort), other states, other loads, other species, or any other apparatus.
• The standing adverse result is reported alongside any positive result, never instead of it: M2_LOGNORMAL out-predicts the two-timescale mixture by ~0.0369 nats event-pooled in the frozen record, and M2 is an adversarial baseline, never the UNI model.

12. Wording

Allowed: "target hypothesis" · "candidate model" · "CI-bound verdict" · "not established" · "duration-only held-out support on the derived_eligible_1_to_8 cohort" · "F-side observational projection" · "retrospective-only" · "transfer required" · "intervention required" · "mechanism discriminator pending" · "G-side design-only until intervention".

Forbidden, mechanically checked by claim_guard.py (use-vs-mention aware). One catalogue entry per line so each line is self-evidently a forbidden-wording listing and not an assertion:

• forbidden wording: "biological parity achieved"
• forbidden wording: "full parity achieved"
• forbidden wording: "active inference demonstrated"
• forbidden wording: "flagellum solved"
• forbidden wording: "full motor solved"
• forbidden wording: "general intelligence"
• forbidden wording: "awareness achieved"
• forbidden wording: "human stack validated"
• forbidden wording: "G proves motor agency"
• forbidden wording: "M2 is the UNI model"
• forbidden wording: "C11 diagnostic proves U4"
• forbidden wording: "mark process prospectively validated on Wadhwa-2022"

13. Reproduction

python -m pytest hierarchical-aif/tests/motor_stack_aif -q
python -c "import sys; sys.path.insert(0,'hierarchical-aif/src'); \
from motor_stack_aif import compare; \
compare.write_result('hierarchical-aif/results/motor_stack_aif/F_SIDE_SCORING_RESULT.json', \
                     compare.run_comparison())"

Deterministic: same inputs produce identical JSON bytes (indent=1, sort_keys=True, LF).

14. NEXT_ACT

Commit this record. Then, and only then, run compare.run_comparison(), write the result, flip this record PENDING -> PROSPECTIVE in the result commit after confirming the ancestry, and score §8 item by item — including the items that missed.