Motor-Stack AIF — Scope Ruling

Gate: H-AIF-G6 · Date: 2026-07-21 · Status: RULING IN FORCE Maps to existing ladder: P1 (implementation); P3 only if scored

1. The ruling

F-side observable projection      : BUILD_AND_SCORE_NOW
G-side biological policy selection: DESIGN_ONLY_UNTIL_INTERVENTION_OR_TRANSFER

The model is not abandoned because one half of it is untestable on passive data. The half that observational data can score is built now; the half that requires actions is fenced with an explicit route to testability, not deleted.

2. Why F-side is buildable now

F_motor is a belief-update objective over observations. It needs only:

• a likelihood for what was recorded (dwell durations, with censoring),
• a hierarchy over the experimental unit (motors),
• an approximate posterior.

All three exist in this dataset. F_motor scores how well a nested generative model explains held-out dwell durations. That is a legitimate observational projection of the stacked architecture, and it is directly comparable to the existing competitors on the frozen split.

F_motor = E_q[ ln q(Theta, eta, z) - ln p(o, z, eta, Theta) ]

Nothing about computing F_motor asserts that the organism performs inference. It is the analyst's objective for fitting a hierarchical model. Keeping that distinction is the whole reason this ruling exists.

3. Why G-side is fenced

G_motor(pi) scores policies over future observations:

G_motor(pi) = KL[ q(o_future|pi) || p(o_future|C_motor) ] + E_q[ H[ p(o_future|z_future,pi) ] ]

A policy must range over actions. In this dataset there are none:

The failure is structural, not sample-size-limited: the action set is empty, so no amount of additional Wadhwa-2022 data would fix it. This is why the fence is a data requirement, not a "collect more of the same" note.

Classification: G_motor biological policy selection = DESIGN_ONLY_UNTIL_INTERVENTION. G may still be used at the ORCHESTRATE level to choose which experiment to run next — that is how B4C02 was selected over the cheaper B4C10 — and that use is a laboratory-process choice, never a claim about the motor.

4. What would make G a biological motor claim

Minimum required data — all three, not any one:

1. A manipulated variable with recorded onset time (load step, PMF step, or electrorotation change), timestamped against dwell boundaries.
2. Paired pre/post observation on the same motors, so the response is within-unit.
3. Enough independent motors under each condition to give a CI-bound verdict on the response. With the current resolution floor of ≈0.042 nats and 19 holdout motors, a between-condition contrast would need substantially more units — the exact number to be derived, not guessed, before any such run.

Absent these, any G-based motor claim is NOT_ESTABLISHED by construction.

5. Lane status under this ruling

6. What the F-side build may and may not move

• May move P1 (equation/implementation) once implemented and tested.
• May inform P3 (held-out predictive) only if scored on the frozen split against the current-design control and the adversarial baselines, with a CI-bound verdict.
• May not move P5, P6, or P8 without additional receipts.
• May not read held-out mark fields (D5 firewall).
• May not treat a favourable F_motor score as evidence that the motor performs inference.

7. Standing frame

Nature supplies the architecture candidate.
The gate supplies the status.
Full biological parity is not a current status. It is the target world defined by the receipts.