Motor-Stack Regime and Dimensionless-Group Check

Type: BUILDER-SUPPORT PROBE (Probe 11). This document moves no P-level, changes no frozen verdict, and creates no claim. It answers one question: what regime does the frozen motor-stack model actually live in, and what would have to be preserved for a transfer to be legitimate?

Split boundary declared: TRAIN_ONLY for every recomputed distributional quantity; HOLDOUT_ALREADY_SPENT_DURATION_ONLY for the published cohort counts (233 events / 19 motors) that are quoted but not recomputed from held-out durations. D5 firewall respected: no nextStateN, direction, or jump value was read at any point in producing this file.

Channel-read disclosure (new, small, and stated rather than hidden): to produce §2 row nominalElectrorotationSpeed I counted the per-motor apparatus label motors[].nominalElectrorotationSpeed for the 19 holdout motors of the frozen cohort. That field is an apparatus setting, not an outcome channel, it is not listed in DATA-CHANNEL-SPEND-LEDGER.md, and its holdout distribution was already referenced by hierarchical-aif/reports/ULTRACODE-TRACK-D-VERIFICATION.md (row for "Experimenter load protocol": "≤ 3 holdout motors at the thin levels"). Every statement I derive from it is marked DESIGN_ONLY and is not evidential.

0. The governing constraint of this probe

Do not transfer biology by shape. Transfer the regime and the ratio.

This repository contains dwell-time behavioural data only. The observed blanket that reaches the frozen model is exactly five fields — {motorId, stateN, durationS, rightCensored, partition} (exhaustively verified in ULTRACODE-TRACK-D-VERIFICATION.md row F3). There is no torque, no viscosity, no proton-motive force, no Reynolds number, no temperature, and no CheY-P channel anywhere in the motor-stack input. Consequently most rows below read NOT-MEASURED. That is the correct and expected outcome of this probe, not a gap in the probe.

NOT-MEASURED in this document is the same status token as NOT_MEASURED elsewhere in the pack: the quantity does not exist in any artifact this model reads, and no value for it may be written.

Two symbol collisions are load-bearing and are flagged wherever they appear:

• tau here is NOT a torque. In hierarchy.py / fit.py, tau is the population standard deviation of the log Weibull shape. It is dimensionless. The thermodynamic work quantity tau * delta_theta (torque times angle) shares the letter and nothing else. They are never the same number and never share units.
• F here is NOT G. The F-side quantity is an observational projection over beliefs. The G-side (expected free energy over policies) is fenced out of the code by a test, because the dataset is passive and the action set is empty — structurally, not for want of sample size.

1. Measurement inventory

Columns are fixed: quantity · measurand · value · unit · scope · source · status · falsifier.

2. Dimensionless groups that ARE computable here

Only four dimensionless groups can be formed from what this repository actually measures. All four are TRAIN_ONLY and all four are properties of this cohort.

2.1 A regime statement that follows directly, and its adverse edge

The frozen model's timescales are fixed only in normalised units. Converting M3's two components back to seconds makes them state-dependent, because y was divided by scale_N:

There is no single pair of seconds-valued timescales in this model. There are eight pairs, tied together by one dimensionless ratio (Π₃ = 11.7787). Any transfer that carries "the motor has a ~10 s and a ~1 s timescale" across to another preparation is transferring a shape, not a regime, and it is not licensed by this evidence.

2.2 Adverse finding — the Fast/Slow labels are inverted at the fitted values

lambda in M3 is a rate (exp(-lf * y), b3-model-competition-runner.py:244-252). The fitted lambdaFast = 0.44485933051063775 is the smaller rate, so the component named "fast" has the longer mean normalised dwell (1/lf = 2.2479015981347104) and the component named "slow" has the shorter one (1/ls = 0.19084408707578607). The optimizer box is lf ∈ [1e-9, 1e4] with no ordering constraint (:424), so nothing enforces the labelling.

Severity: naming/resonance defect, not a numerical defect. No score, contrast, verdict, or gate depends on the label — M3's density is symmetric under relabelling and the frozen NLPD 3.4343333331 is unaffected. The hazard is purely in reading: a report that says "the fast component carries weight 0.393" is describing the long-dwell component. Correct minimal reading: the mixture has a long-dwell component with weight w = 0.3933559993214189 and mean 2.2479 in normalised units, and a short-dwell component with weight 1 - w = 0.6066440006785811 and mean 1/ls = 0.19084408707578607. Recommended correction is documentation only — audits/** is frozen and must not be edited.

3. What would have to be matched for a transfer claim to be legitimate

A transfer of this model to another preparation, species, or apparatus is legitimate only if the following are matched or explicitly bounded. Anything unmatched is an extrapolation and must be labelled as one.

Consequence. Of the ten preconditions, three (T1, T2/T3 partially, T9/T10) are measurable in this repository and six are NOT-MEASURED outright. A transfer claim is therefore transfer required / intervention required — the missing groups are missing data, not missing analysis, and no amount of further modelling on Wadhwa-2022 supplies them. This is the same wall the P-ladder already records: P4/P5/P7 cannot be closed by any modelling in this repository.

4. Validity-domain map

Classification vocabulary per CLAUDE.md: supported · tentatively supported · contradicted · unidentifiable · unobserved · extrapolation-only.

Count: of 15 domain dimensions, 2 are supported, 1 is tentatively supported, 2 are unidentifiable, 1 is mixed supported/extrapolation-only, 8 are unobserved, 0 are contradicted. That distribution — mostly unobserved — is the honest regime portrait of this model.

5. What this probe did NOT do

• NOT_RUN — COMPUTE_BUDGET: no bootstrap, refit, or simulation was executed. Two corrected-full runs (B4C11, B4C01) are in flight and no compute was spent that could contend with them. Their live progress counters were not read and are not cited anywhere in this document.
• NOT_CHECKED — would require holdout access: the held-out normalised-y range, the held-out per-state dwell means, and any held-out recomputation of Π₁/Π₂. Every distributional number here is TRAIN_ONLY.
• NOT-MEASURED: every physical measurand in §1 marked as such. No value was estimated, imputed, borrowed from literature, or carried across from another study. The Antani-2021 torque and CheY channels exist in this repository and were deliberately not used as substitutes — they belong to a different study, a different assay, and different motors.
• No threshold was invented. The only thresholds used are frozen ones (the ≈0.042 nat resolution floor; the B4C10 U2/U3/U4 firing rules, quoted but not re-evaluated). Every design statement derived from the load label is marked DESIGN_ONLY and is not evidential.

6. Falsifiers for this document itself

1. Recomputing scale_N from the frozen TRAIN partition and getting different IEEE doubles.
2. Finding any torque, viscosity, temperature, PMF, or Reynolds field reachable by the motor-stack likelihood. (Exhaustive field extraction says the interface is five fields; a sixth would falsify §0.)
3. Finding an ordering constraint on lf that makes the Fast/Slow labels of §2.2 correct.
4. Recovering t_step from the raw .mat c14de12c…, which would move the load dimension out of unidentifiable — retrospectively only, never prospectively.

NEXT_ACT = Add a documentation-only correction note to hierarchical-aif/docs/ recording the M3 lambdaFast/lambdaSlow label inversion found in §2.2 (frozen audits/** stays untouched), and register T1–T10 as the transfer preconditions checklist in the P4 transfer lane so no future transfer proposal can skip a NOT-MEASURED group.