Cross-study biological parity program

Verdict

The expanded UNI-FLAGELLUM model expresses substantially more observed biology, but full and complete parity is false. The frozen protocol has 16 conjunctive gates: 8 PASS, 3 FAIL, 2 NOT_ESTABLISHED, and 3 BLOCKED_EXTERNAL. No score averaging can turn a missing required domain into parity.

The strongest new positive result is not a claim of one universal equation. It is a traceable modular model spanning stator dwell kinetics, rotation-gated assembly, stator/CheY association, torque-conditioned switching, a non-equilibrium switching generator, occupancy statistics, and whole-cell propulsion. The strongest new falsifier is that the exact 13-site lattice gives incompatible interaction strengths when fitted to full occupancy distributions versus occupancy moments.

This release is deterministic, CPU-only, and contains no LLM or GPU in the runtime or analysis path.

Evidence population and independence

The corpus contains 11 explicitly attributed studies across five observation scales. Four study families provide direct primary artifacts. The conservative lower bound is 409 independent motors or cells:

The lower bound deliberately avoids pseudoreplication. Ito's source workbook contains 191,920 rotation samples and 159,800 stator-occupancy samples, but those 351,720 time samples are not counted as independent organisms. Multiple measurements of one motor remain one experimental unit.

The separately cached Ito raw-trace archive is 4,085,227,742 bytes. A full streaming verification matched MD5 d42879e66142ff7190f256f4276db111, parsed 505 ZIP entries, and found no member CRC failure. The normal test suite checks the committed verification ledger instead of rehashing 4.09 GB. npm run cross-study:verify-raw repeats the deep archive audit when the external cache is present.

Evidence classes

The protocol was frozen after inspecting published sources but before coded gate execution. It is therefore a post-publication reproduction and stress test, not a blind preregistration or independent confirmation.

Model organism: modules, boundaries, and equations

The model is one auditable organism in the engineering sense: every module has a declared biological domain, observation operator, units, evidence tier, and claim boundary. It is not one undifferentiated parameter vector. Parameters cross a study boundary only when strains, interventions, loads, units, and measurement operators are commensurate.

M_DLT: post-perturbation stator dwell kinetics

For stator occupancy N, the source-derived first-passage survival is a finite mixture:

S_N(t) = sum_j a_j(N) exp[-r_j(N)t]
r_j(N) = k_+(N) + N sigma_- + j(sigma_+ - sigma_-)
P_+(t|N) = k_+(N) S_N(t)
P_-(t|N) = -dS_N(t)/dt - P_+(t|N)

Observed event direction, dwell duration, censoring, and motor identity define the likelihood. This module remains partially identified; its hidden timescales are not molecular identities.

M_ROTATION_GATE: signed rotation-dependent stator binding

b(omega) = b0 + a_plus max(omega, 0) + a_minus max(-omega, 0)

Weighted fits use the source-reported standard errors. Prediction is evaluated by leaving out a complete speed bin, never an individual time point. The 20-Hz-bin fit gives:

b0      = 0.1544997646 s^-1
a_plus  = 0.0176841082 s^-1 Hz^-1
a_minus = 0.0319566693 s^-1 Hz^-1
LOO RMSE = 0.8293641
constant-baseline LOO RMSE = 1.4544735
relative improvement = 42.9784%

The criterion also passes at 5-Hz and 10-Hz bin widths. This supports a signed speed-rate relationship in the source aggregate. It does not make speed bins independent motors, and it does not identify a molecular sensor.

M_LATTICE13: exact finite occupancy lattice

For binary occupancy phi_i on a periodic lattice of length 13:

P(phi | J, mu) = exp[J sum_i phi_i phi_(i+1) + mu sum_i phi_i] / Z(J, mu)
P(N | J, mu) = sum_{phi: sum phi_i=N} P(phi | J, mu)

The implementation enumerates all 2^13 = 8192 states exactly. It jointly fits three bead/load distributions with a shared J and load-specific mu. It also independently fits the reported occupancy standard deviations, which reproduces the published moment estimate.

This is a decisive internal disagreement:

full-distribution J = 0.2106346292
moment-fit J         = 1.1875632937
reported-SE-weighted moment J = 1.0705756792
absolute cross-statistic difference = 0.9769286644
full-distribution probability RMSE = 0.0344230431
Delta AIC in favor of J>0 model = -1.4588058431

The probability RMSE is small, but the extra cooperative parameter is not favored over J=0, and the preregistered range/AIC criteria fail. This does not erase the measured occupancy distributions. It shows that a single molecular interaction energy is not identified consistently by these aggregate statistics under this lattice model.

M_GMC: non-equilibrium switching generator

The CPU port uses a continuous-time Markov generator over engaged and unengaged conformational counts. Local torque exponentially modulates engaged subunit switching rates; rotor motion exchanges engaged and unengaged subunits. For generator Q and stationary distribution pi:

Q_ij >= 0 for i != j
sum_i Q_ij = 0
Q pi = 0
sum_i pi_i = 1

At the frozen source condition, the independent port has 175 states, maximum column-sum error 8.53e-14, stationary residual 9.55e-15, and engaged-state marginal L1 error 0.0176053 against the authors' source output. This verifies the ported generator and stationary calculation, not experimental truth.

Against attributed Yuan 2009 switching observations inside the source model's speed support, 7 of 11 points in each switching direction fall within the digitized asymmetric error bar. Mean absolute standardized residuals are 0.7876 for CW-to-CCW and 0.8760 for CCW-to-CW. This passes the frozen source reproduction gate but is not held-out cross-laboratory parameter transfer.

M_RFT: whole-cell propulsion

Published resistive-force-theory predictions are compared with independently reported cell speed at matched flagellar count. The ten-point source comparison has RMSE 1.7274107 um/s, versus 2.0537120 um/s for a constant-mean baseline, a 15.8884% improvement. The measured motor-speed slope is -0.5112 Hz/flagellum, with 95% interval [-2.1945, 1.1721], which includes zero. This supports the declared source-level propulsion gate, not a universal strain- and medium-independent law.

Additional observed couplings

Antani motor/cell-level fluorescence gives a MotAB-associated CheY signal difference of 1.13324, Hedges g=0.7774, Welch p=0.000817, and a 5,000 motor-bootstrap 95% interval [0.56641, 1.82915]. The required FliM and rotating-versus-stalled controls are retained and nonsignificant (p=0.6087 and p=0.5678). This supports association/coupling, not an unrestricted causal chain.

Across nine Antani torque conditions, torque correlations are rho=1.0000 for CW-to-CCW and rho=0.9333 for CCW-to-CW, while CW-bias coefficient of variation is 0.1000. These are aggregate condition-level results because a shared-cell covariance artifact is unavailable.

Gate ledger

Structural and CAD truth boundary

Five model-state mappings are explicit in experiments/structural-state-map.v1.json. Two conflicts remain:

1. The 13-site occupancy lattice is an assay-specific statistical model, not established E. coli molecular slot geometry.
2. DLT weak/strong or age-mixture states are phenomenological until a separate molecular measurement identifies them.

The physical export is therefore CAD-ready as a transparent, manipulable UNI mathematical model. It is not a literal molecular replica of the flagellar motor. Only dimensions and multiplicities with direct structural provenance may drive literal molecular geometry; phenomenological states must remain visually labeled mathematical layers.

What a complete model still requires

The next discovery cycle is constrained by evidence, not by adding equations:

1. Commensurate raw transfer: identify two laboratories with compatible strain, load, perturbation, calibration, and observation operator; freeze parameters on lab A and score untouched lab B against declared baselines.
2. Latent-state identification: collect a simultaneous molecular reporter and rotation/stator-occupancy trace to test whether DLT mixture states map to specific conformations or binding states.
3. Distribution-level cooperativity: acquire motor-resolved occupancy distributions and repeated load conditions so cell heterogeneity is not absorbed into J; compare lattice, independent-site, mixture, and kinetic alternatives with held-out motors.
4. Active-Inference discrimination: preregister an intervention for which an Active-Inference-specific policy/preference model and matched kinetic, feedback-control, and non-equilibrium models make different quantitative predictions. Bayesian fitting alone is not such evidence.
5. Live prospective run: timestamp raw voltage/position, calibration, uncertainty, model version, and prediction before revealing the outcome.
6. Independent replication: transfer the frozen protocol, including exclusions and failure rules, to an independent laboratory and publish all raw files and calibration.
7. Physical validation: fabricate the abstract UNI model, inspect tolerances/backlash, calibrate sensors, and bind a hand-driven mechanical trajectory to the same on-screen math ledger.

No software-only action can honestly complete items 5-7. They remain executable experimental contracts rather than simulated passes.

Reproduction

The large raw archive is external and ignored by Git. With it present, repeat the deep integrity check once; the other commands are routine and CPU-only.

npm run cross-study:verify-raw
npm run cross-study:ingest
npm run cross-study:run
npm run cross-study:verify
npm run science:verify
npm test
npm run lint

cross-study:verify is a separately implemented JavaScript checker for the Ito leave-one-bin-out error, exact-lattice SSE and optimized J=0 baseline, GMC source comparisons, RFT error, and audit hashes. Running the Python runner twice must leave report and audit bytes unchanged.

Machine-readable execution artifacts:

• experiments/cross-study-preregistration.v1.json
• experiments/data/cross-study-motor-evidence.json
• experiments/structural-state-map.v1.json
• experiments/results/ito-raw-archive-verification.json
• experiments/results/cross-study-parity-report.json
• experiments/results/cross-study-parity-audit.json
• scripts/ingest-cross-study-evidence.py
• scripts/run-cross-study-parity.py
• scripts/independent-cross-study-check.mjs

The public copies under public/ are byte-compared in tests where applicable, so the rendered Verum ledger and downloadable evidence cannot silently diverge from the experiment results.