Next-Gate Plan — Transfer and Intervention

Date: 2026-07-22 · Gate targets: P4 transfer, P5 intervention, P6 mechanism · Status: DESIGN-ONLY. This plan proposes receipts; it runs nothing, moves no P-level, and asserts no claim.

Every threshold in this plan is DESIGN_ONLY unless it is a frozen criterion. The dataset in hand is passive E. coli dwell-time behaviour (Wadhwa 2022); everything below requires data this repository does not have.

0. Why a next gate at all — the binding constraint, measured

The corrected phase established, from two independent directions, that the binding constraint is data, not model math:

• B4C01 (assay side): the 19-motor motor-equal design is nesting-blind — it cannot separate nested near-equivalent dwell families even under correct specification.
• F-side + D10 (model side): the candidate is NOT_ESTABLISHED against every serious adversary and the CI rule resolves numerically-null differences.

Re-scoring the same duration data cannot move P4–P8. The next gain requires an experiment that makes structurally distinct hypotheses disagree. Strong inference: keep them all alive until a predeclared experiment can exclude one or more.

1. What independent dataset would discharge P4 (transfer)?

Requirement: a second held-out cohort of single-motor dwell-time series, independent of Wadhwa 2022, with:

• enough independent MOTORS that a motor-equal contrast can resolve a real difference — B4C01/the power atlas indicate 19 is nesting-blind; a DESIGN_ONLY target is the motor count at which a 0.042-nat contrast resolves at ≥80% (from the synthetic power atlas, not yet measured on real data — NOT_MEASURED);
• a predeclared sha256_mod5-style split committed before any scoring (D9 discipline);
• the same observable (dwell duration, state, censor flag) so the frozen scoring rule transfers without a units change.

What it would license: a P4 transfer statement — a scoped result holding on a second cohort. What it would not: any mechanism or parity claim; transfer is necessary for P8, not sufficient. Candidate sources (design-only, unverified): other published bead-assay / fluorescent-switch motor datasets; a re-analysis partner. Species discipline: keep E. coli behavioural evidence separate from Salmonella/Bacillus structural evidence — a structural dataset does not transfer a behavioural claim.

2. What perturbation would make G-side policy selection testable (P5)?

The fence, restated: the current dataset is passive and the action set is EMPTY — structural, not sample-size. expected_free_energy does not exist in the package and a test enforces its absence. G stays DESIGN_ONLY_UNTIL_INTERVENTION_OR_TRANSFER. This section proposes what data would change that, not code to add now.

Candidate perturbations, each needing a manipulated variable with recorded onset time, paired pre/post on the same motors, and enough independent motors per condition for a CI-bound verdict (minimum N NOT_MEASURED — derive from a real-data power atlas once one exists):

The core discriminator this must answer (§3).

3. What mechanism discriminator would separate AIF-motor-stack from M2 / M4 / M7?

This is the crux and it is currently unanswered — stated honestly rather than filled.

The models to separate: AIF motor-stack policy selection · M2/lognormal shape · M4/M7 hierarchy · three-timescale kinetics · censoring artifact. On passive duration data they are near-equivalent (F-side ≈ M7 to 2.5e-7; M2 leads; B4C01 shows the assay is nesting-blind). So no duration-only discriminator exists — that is the measured finding, not a gap to paper over.

A discriminator must therefore live in a channel the models predict differently:

• AIF policy selection predicts a response to intervention that a fixed-rate model does not — a change in the policy (dwell-state transition structure) following a perturbation onset, beyond the change a passive rate model predicts from the altered load/PMF alone. Naming a concrete, pre-registerable such prediction is the open design task. If it cannot be named, G stays fenced and unfalsifiable here — which is itself the correct, reportable status.
• Three-timescale kinetics vs shape is separable with a mark/transition channel (the dwell sequence, not just durations) — which routes through §4.
• Censoring artifact is separable with a censored-inclusive cohort — the F-side censoring branch is correct-by-test but unexercised on real data (the frozen cohort excludes censored events).

Honest status: no single discriminator yet separates all five on obtainable data. The plan is to build the ones that are namable (censoring-inclusive scoring; a transition-channel likelihood after D5/D6 repair) and to flag the AIF-vs-passive discriminator as requiring an intervention design that does not yet have a concrete pre-registerable prediction.

4. What mark-process repair is required after D5 / D6?

Detailed in protocols/MARK-PROCESS-TRANSFER-RESCUE-PROTOCOL.md. Summary of what is owed:

• The Wadhwa holdout mark channel is retrospective-only (D5, burned) — a new prospective split or an independent dataset is required for any prospective mark claim.
• The nextStateN range defect (D6) must be repaired at ingest (2 holdout events carry nextStateN = -1; 15–17% of marks leave {1..8}) via raw-archive re-derivation before any closed-chain likelihood is valid. marks.py already offers strict/quarantine/retain policies; no silent-drop policy exists.
• Smoothing must be predeclared — the ledger records that the smoothing constant currently flips the sign of the result, so it must be committed before any new mark analysis.

5. What result would KILL the biological-parity target hypothesis?

H_PARITY (the target world, kept alive as a hypothesis) would be falsified / fatally weakened by:

• a simple adversary winning by a CI-bound, MATERIAL margin on an independent transfer cohort — e.g. M2/lognormal beating the motor-stack candidate above the resolution floor on a second dataset (not merely NOT_ESTABLISHED, and not sub-floor per D10);
• an intervention showing dwell-state reorganisation fully explained by a passive rate change — no policy-like signal beyond the mechanical consequence of the perturbation;
• parameter-recovery failure on real transfer data — the fitters biased on genuine out-of-distribution motors (distinct from B4C01's benign synthetic recovery);
• irreparable mark-process invalidity — the impossible marks proving a real ingest/model incompatibility rather than a fixable range-check.

6. What result would STRENGTHEN it?

• A risky prospective prediction whose prospectivity is decided by the commit graph, not by prose (D9) — its record committed ahead of the observation, confirmed on an independent cohort — the only kind of evidence that moves P4/P7 and that this programme's discipline is built to produce.
• An intervention discriminator where AIF policy selection predicts a response M2/M4/M7 do not, and the observed response matches — the only route to a non-fenced P5/P6.
• The censoring-inclusive cohort showing the F-side censoring branch carries real predictive advantage — the one currently-untested distinctiveness of the F-side model.
• A transition-channel likelihood (post D5/D6 repair) separating three-timescale kinetics from shape on valid mark data.

None of these is achievable on the data in hand. That is the honest boundary of this phase, and reporting it is a legitimate scientific outcome, not an incomplete task.

7. Priority order (expected information gain × obtainability)

1. Censoring-inclusive cohort + rescore — highest obtainability (needs only a re-derived cohort retaining right-censored events), directly tests the F-side model's one untested distinctiveness, and discriminates the censoring-artifact hypothesis. Pre-register first.
2. Independent transfer cohort — moves P4, directly relaxes B4C01's nesting-blindness. External data; highest scientific value.
3. Mark-process repair (D6 raw archive) + transition-channel likelihood — separates kinetics from shape; blocked on external archive.
4. Intervention design — moves P5/G-side; blocked on the open discriminator question of §3 and on wet-lab data.

NEXT_ACT = prepare the commit (no push); present git status, files, tests, guards, hashes, and the proposed commit command to the principal