Nature-Inferred Motor Design Map

Status: BUILDER-SUPPORT DESIGN MAP. Raises no P-level, changes no frozen verdict, creates no claim. It answers one question per row: which design element came from a biological observation, and what does the receipt actually say about it?

0. The fence — read this before the table

Nature motivates the candidate architecture; gates decide status.

A nature citation is never a UNI gate. Nature is a search heuristic, not a proof surface. Citing a flagellar-motor observation tells you where it was reasonable to look for a modelling principle. It tells you nothing about whether the resulting software predicts held-out dwell times, and it can move no gate by itself.

Three separate things are kept apart in every row below and must never be collapsed:

The nature observation column is the inspiration column. Reading it as evidence for this cohort is the exact category error this document exists to prevent. If a row's current test cell says NO_CURRENT_TEST, then no amount of biological citation in that same row upgrades it.

Species discipline (contract-binding)

This cohort is E. coli behavioural evidence: Wadhwa 2022, stator-occupancy dwell events, cohort derived_eligible_1_to_8, 80 train / 19 holdout motors, 793 / 233 events.

Structural facts from other species appear in this repository — Salmonella rod/hook/basal-body (PDB 7E82), Salmonella switching visualisation (Singh et al. 2024, DOI 10.1038/s41564-024-01674-1), Bacillus subtilis MotA5MotB2 homologous geometry (PDB 6YSL), catalogued in docs/LIVING-SCIENCE-WALKTHROUGH.md. Wherever such a fact is named below it is labelled STRUCTURAL / <species> and carries this rider:

A Salmonella or Bacillus structural fact is not behavioural evidence and is not evidence for this E. coli cohort. These did not come from one measured specimen, and no row below may be read as though they did.

Status vocabulary used

1. The map

Columns are terse by design; §2 expands each row with its receipt. Wide table — scroll right.

2. Row detail — the receipt behind each status

R1 · nested motor hierarchy Lmotor-5..Lmotor-0

The architecture as built: Lmotor-5 population prior (mu, tau) → Lmotor-4 per-motor latent eta_m (integrated) → Lmotor-3 occupancy state (frozen scale_N) → Lmotor-2 kinetic mode (not instantiated, R10) → Lmotor-1 hazard/survival → Lmotor-0 observed blanket.

The clean ablation of the hierarchy is not M0. It is M1_WEIBULL — the pooled, single-shape tau -> 0 limit of the same likelihood family. Verbatim from F_SIDE_MOTOR_STACK_SCORING_RESULT.json contrasts.M1_WEIBULL: pointEstimate 0.0006145100976836924, interval [-0.010880659688579656, 0.011303987651464963], halfWidth 0.01109232367002231 (percentile, intervalType: "percentile" — D7), resamplingUnit: "MOTOR", nRep 2000, seed 20260717, verdict: "NOT_ESTABLISHED", atOrBelowResolutionFloor: true.

The nesting currently buys nothing resolvable on 19 holdout motors. This is an adverse result for the architecture and is reported as the headline of this row, not as a footnote. The artifact's own interpretation field says it: this is not equivalence and not "no difference".

OPEN: D1 — the C11 cluster-collapse defect invalidates the previous U4_OK reading; the corrected full run is in flight. Nothing in this document may be read from that run's live counters, which are telemetry, not results.

R2 · censoring — the honest gap

test_censoring_likelihood.py proves the branch is mathematically right, including a mutation test showing that mislabelling one event moves the total log-likelihood, and a no-floor test. But the frozen cohort derived_eligible_1_to_8 excludes right-censored events entirely. So the branch is correct-by-test and unexercised by any scoring run. That is a coverage gap in the implemented likelihood, not a property of any result.

R3 · hazard/survival timing

The only element in this map whose biological-flavoured reading is supported by a resolved held-out contrast. Verbatim from contrasts.M0_EXPONENTIAL: pointEstimate 0.11529094279542929, interval [0.013039958895400987, 0.22749707235626093], halfWidth 0.10722855673042997 (percentile), verdict: "RESOLVED_ABOVE", atOrBelowResolutionFloor: false. The point estimate sits above the ≈0.042 nat resolution floor, so the effect is material, not merely resolved — note that the interval's lower end 0.0130… does not, so materiality here is a statement about the point estimate, not a CI-bound guarantee of materiality. Scope discipline: this supports non-exponential dwell timing on this cohort. Predictive superiority is never promoted to mechanism.

R4 · motor-specific heterogeneity and D10

fitted.F_MOTOR_STACK: mu = -0.41607215987582913, tau = 0.18372082607308418, k_population_median = 0.659632669755436, trainNLL = 575.6701064153622, converged true, tauAtBoundary false, nfev 165, nit 83. Against M7_HIERARCHICAL_MOTOR frozen tau = 0.18372185667134974 and k = 0.6596322379287862. They agree to ~7 significant figures. From M4_M6_M7_PER_MOTOR_CONTRASTS_RESULT.json contrasts.M7_HIERARCHICAL_MOTOR: pointEstimate 2.5069835851709854e-07, interval [1.6044512352276248e-07, 3.3746883763884875e-07], verdict "RESOLVED_ABOVE", scientificReading.classification "SCIENTIFICALLY_NULL", floorToEffectRatio 167532.01037467283 against resolutionFloorNats 0.042. That contrast is the numerical shadow of two implementations of the same model — which is exactly why D10 was raised and closed by adding interpretation, never by moving a threshold.

R5 · state conditioning — structurally undiscriminable here

Every model on the leaderboard, candidate and adversary alike, consumes the same frozen per-state scale_N. It is a shared change of variable. Its +log(scale_N) Jacobian is load-bearing for units (mutation test halts the oracle when it is dropped) but it cancels in every contrast. No receipt in this repository discriminates state-conditioned from state-pooled dynamics, and none can until a state-pooled ablation is built and scored on the same split.

R6 · the blanket boundary

Typed, immutable, mode-gated, and defended by 27 test functions across two files (14 + 13) including detector positive and negative controls. What it establishes is that the software cannot cross its declared boundary. It establishes nothing about where a biological Markov blanket sits, and this row must not be cited as though it did.

R7 · switch/dwell structure — the burned lane

Two independent defects, both preserved:

• D5 — the held-out mark channel was read by a read-only audit brief with no split boundary. Irreversible. One study, no second holdout, unrepairable in this dataset.
• D6 — impossible marks (nextStateN = -1, 2 holdout events), 5 holdout events with zero training support, and an open alphabet (15.1% train / 16.7% holdout marks leave {1..8}). The frozen dataset is not edited; marks.py forces an explicit policy decision instead.

test_nextstate_range_check.py and test_no_holdout_mark_read.py keep both quarantines live.

R8 · mean-one normalisation

Note the earned lesson embedded here: the unit-mean check uses adaptive quadrature precisely because a uniform-grid mean-one integral can miss a y^(a-1) singularity while the analytic node-weight sum is exact. The exact guarantee and the numerical check are kept as separate objects.

R9 · motor as experimental unit

The D10 interaction matters for design: a paired motor-cluster bootstrap resamples motors, so it resolves a difference of any magnitude whose sign is consistent across motors. The ≈0.042 nat floor predicts what is scientifically material; it does not predict what the bootstrap will call. Both numbers are needed and they answer different questions.

R10 / R11 · the two deliberate absences

Lmotor-2 is absent because it would be unidentifiable. G-side is absent because there is no action set. These are different reasons and must not be merged: the first is a resolution argument that more motors could in principle overturn; the second is structural and more passive data cannot touch it.

3. Where this document disagrees with the existing rationale

docs/NATURE-INFERRED-MOTOR-STACK-DESIGN-RATIONALE.md is the parent document and this map is consistent with it, with three deliberate divergences recorded here rather than made silently:

1. Rationale §3 lists "Hierarchy" as an extracted principle that "lands" at Lmotor-5/4. That is true as implementation. This map adds the receipt the parent predates: the hierarchy level is TESTED_NOT_ESTABLISHED against its own tau -> 0 limit (M1, +0.000615 [-0.010881, +0.011304]). The parent should not be read as implying the principle is supported by evidence; it is supported by motivation.
2. Rationale §3 lists "State conditioning" as landing at "per-state scale normalisation". This map records that the normalisation is shared by every competing model, therefore structurally undiscriminable in any current contrast (R5). The parent does not say this.
3. Rationale §9 falsifier 2 says "Falsifier 2 is live right now — B4C02 is running." That statement is now stale: B4C02 LANDED (0633988d…). Its outcome was gensWithM2overM3 = 1 of 3 → GENERATOR-SPECIFIC, so the frozen prediction of GENERATOR-ROBUST_ADVERSE was REFUTED — the adverse M2-over-M3 result is not a generic heavy-tailed shape artifact. The parent document's §9 text should be updated when someone next edits it; this map does not edit it.

No other disagreement. The parent's §10 forbidden readings are adopted here verbatim in spirit and extended by §0 above.

4. Gaps this map exposes (design observations, not verdicts)

DESIGN_ONLY thresholds introduced by this document: none. Every numeric threshold quoted (≈0.042 nat floor, 5/793 = 0.006305, 0.25, 2.0 decades) is a frozen criterion quoted from an existing artifact, not a new one.

5. Forbidden readings of this map

• that a filled nature observation cell is evidence for the row's software element;
• that a row with NO_CURRENT_TEST is neutral or presumed fine — it is unmeasured;
• that TESTED_NOT_ESTABLISHED means the two models are equivalent. It does not. With 19 holdout motors most contrasts are inconclusive; underpowered is not equivalence;
• that Salmonella or Bacillus structural evidence says anything about this E. coli behavioural cohort, or that any of them came from one measured specimen;
• that computing F implies the motor performs inference. The B3 result's own notEstablished disclaims that reading and this map does not license it;
• that this document moved any P-level. It moved none. P8 remains FULL_PARITY = false and the first unsatisfied level remains P4 transfer.

NEXT_ACT = build the R5 state-pooled ablation (gap G1) as a DESIGN_ONLY-free, receipt-bearing contrast — a single-global-scale variant scored on the frozen derived_eligible_1_to_8 split under the same paired motor-cluster bootstrap — and pre-register its prediction record BEFORE running it, so state conditioning stops being IMPLEMENTED_UNTESTED; do not launch it while B4C11 and B4C01 are in flight