Nature-Inferred Motor-Stack Design Rationale

Status: DESIGN RATIONALE. Raises no P-level by itself.

Nature supplies the architecture candidate. The gate supplies the status.

Biology may motivate what we build. Only a receipt moves a gate. Never write "nature proves parity."

1. Function

Bacterial flagellar-motor behaviour: stator occupancy and remodelling, transitions between occupancy states, dwell timing in each state, switching, and load adaptation.

2. Constraint regime

Molecular motor; stochastic transitions; limited observations (1349 events, 99 motors, one study); strong motor-level heterogeneity; right-censored dwells (109, one terminal per motor trace); no recorded interventions.

3. Extracted principles (what we take)

4. What is FORM — deliberately not copied

Stator stoichiometry, torque-speed curves, protein structure, PMF biophysics, ion flux. These are biological embodiment details. Reproducing them in software would be imitation, not evidence, and would not make any gate move.

5. What is PRINCIPLE — implemented

Nested inference over an exchangeable unit; survival-aware likelihood; per-unit heterogeneity integrated rather than freely estimated; state-conditioned dynamics; a strict observation boundary.

6. What can be tested NOW

• duration likelihood on held-out motors (durationS was prospectively spent — legitimate)
• censoring correctness
• motor-level hierarchy
• state conditioning
• comparison against control and adversarial baselines under a CI-bound verdict

7. What CANNOT be tested now

8. Deliberate simplicity

The constrained F-side model has 2 free parameters. Per-motor latents are integrated by 33-node Gauss-Hermite quadrature rather than estimated, so parameter count does not grow with motor count. A model with 80 free per-motor shapes would fit 80 numbers to a median of 7 events each and could not be honestly scored.

Adding Lmotor-2 policies would add unidentifiable capacity, not testability. That is why the level is specified in the architecture and deliberately not instantiated. Against a corrected resolution floor of ≈0.042 nats, flexibility that cannot be resolved is not neutral — it is a way to look complete while being untestable.

9. Falsifiers

The design is wrong, or at least unsupported, if:

1. adversarial dwell baselines (M0/M1/M2/M5/M8) win or tie under corrected CI-bound tests;
2. corrected B4C02 shows the adverse M2-over-M3 result reproduces under misspecified generators — i.e. it is a heavy-tailed shape artifact, not mechanism-bearing;
3. corrected B4C11 fails to support the motor-level hierarchy;
4. the model proves unidentifiable at the achievable resolution;
5. independent transfer fails.

Falsifier 2 is live right now — B4C02 is running, and its frozen prediction expects GENERATOR-ROBUST_ADVERSE, which if confirmed weakens the mechanistic reading of M3's loss. The design was committed to that test before its outcome was known.

10. Forbidden readings

• that biological correspondence constitutes parity evidence
• that architectural inspiration substitutes for a receipt
• that the motor stack demonstrates anything about the human L0..L12 stack
• that computing F implies the motor performs inference