Scientific parity gates

Cross-study extension: This document records the original Wadhwa-focused gate cycle. The broader, later protocol is documented in CROSS-STUDY-PARITY.md. It adds 11 attributed studies, at least 409 independent motors/cells, rotation/load/switching/propulsion layers, and 16 new gates. Evidence relevant to the original G08/G09 now exists in assay-specific modules, but unit-safe cross-laboratory parameter transfer is still not established and must not be inferred by merging unlike experiments.

Verdict

The current release has partial computational parity only. Four of seven executable computational gates pass. Three fail. Five further gates require new instrument, biological, laboratory, or physical-print evidence; a biological Active-Inference identity is not established.

This is the intended behavior of a science gate: a negative result remains a negative result in the JSON, tests, documentation, and rendered laboratory.

What was added

The previous observed experiment fitted generic normalized duration families. It did not implement the mechanism analyzed by Wadhwa et al. The parity layer now implements the source paper's D–L–T first-passage reduction:

S_N(t) = Σ_j a_j exp(-r_j t)
r_j = k_+(N) + Nσ_- + j(σ_+ - σ_-)

P_+(t|N) = k_+(N) S_N(t)
P_-(t|N) = -dS_N(t)/dt - P_+(t|N)

For N=1, a=[1-c1,c1]; for N=2, the weights are the convolution of [1-c1,c1] and [1-c2,c2]; and for N≥3, a third factor using c3 is included. The weights are normalized and nonnegative under the fitted constraints.

An observed on transition contributes log P_+(t|N), an off transition contributes log P_-(t|N), and a right-censored interval contributes log S_N(c). All events from one motor stay in one partition. Parameters are fitted only on training motors.

The paper handles the short-lived H state separately by a threshold classifier. The present likelihood therefore cannot honestly be called a unified D–L–T–H likelihood. H-state reproduction remains a separate gate.

Gate results

Public-artifact discrepancy

The article reports a moment fit near c1=0.30, c2=0.12, c3=0.06, σ+=0.19 s^-1, and σ-≤0.0005 s^-1. Its uncertainty notation is the parameter displacement producing a 50% loss increase, not a confidence interval.

The bundled file fitting_parameters.txt instead contains c1=0.613638, c2=0.308125, c3=0.162786, σ+=0.095787 s^-1, and σ-=10^-8 s^-1. Running the published code equations with that vector produces a maximum relative Figure 3 mean-dwell discrepancy of 3.767, a maximum absolute f+ discrepancy of 0.443, and a maximum absolute normalized-variance discrepancy of 3.862 against the article's source-data workbook.

There is a second internal discrepancy. The paper's equation says N=0 has a single exponential survival, which requires normalized variance V=1. The source workbook reports theory V(0)=1.5920577617. The public Python function applies its c1 mixture branch to N=0 for mean and variance while separately forcing f+(0)=1. The ledger records this conflict; it does not choose a silent correction.

This finding concerns reproducibility of the public theory artifacts. It does not erase or refute the observed single-motor records.

Mechanistic fit and identifiability

The training-only censored maximum-likelihood fit gives:

σ+ = 0.1173812392 s^-1
σ- = 0.0005383373 s^-1
c1 = 0.4623986758
c2 = 0.0004212689
c3 = 0.0000010723

The collapse of c2 and c3 toward zero matters. Under this split, extraction, and likelihood, three arrival-age coefficients are not practically identified. The synthetic recovery suite reached the frozen acceptance bounds in two of three seeds; the third missed c2 by 0.1384 when the limit was 0.12.

This is not a reason to remove the failed gate. It is evidence that more data, a stronger measurement model, or a simpler parameterization is needed before the coefficients can be assigned stable biological meaning.

Held-out prediction

The mechanistic model and a state-specific memoryless competing-risk baseline were fitted on the same training intervals. Every eligible held-out interval, including its censoring indicator and observed exit direction, was then scored.

D–L–T mean log score:        -3.83808 nat/interval
memoryless mean log score:   -3.89634 nat/interval
difference:                   0.05826 nat/interval
motor-cluster 95% interval:  [-0.01580, 0.13264]

The point estimate favors D–L–T, but the interval includes zero. Therefore the frozen predictive gate fails. It is not reported as “almost proved.”

Reproduction and audit

python -m pip install -r requirements-experiments.txt
npm run science:run
npm run science:verify
npm test

science:run performs the CPU-only fit, 2,000 motor-cluster bootstraps, and three deterministic parameter-recovery experiments. science:verify is an independent JavaScript implementation of the first-passage equations and held-out likelihood. Running science:run twice must produce the same report and audit hashes.

Machine-readable artifacts:

• experiments/source-parity-reference.json
• experiments/results/science-gates-report.json
• experiments/results/science-gates-audit.json
• lib/source-first-passage.js
• scripts/run-science-gates.py
• scripts/independent-science-check.mjs

Work required for biological parity

1. Obtain a tagged final source parameter artifact or author clarification for G03 while retaining the original mismatch in provenance.
2. Freeze and independently implement the H-well classifier for G07.
3. Acquire raw motor-identified multi-load occupancy, torque, and switching observations for G08 and G09.
4. Prospectively commit predictions before a calibrated live run for G11.
5. Obtain an independent laboratory replication for G12.
6. Print, instrument, measure, and safety-review the physical UNI mechanism for G13.

Until those gates are actually executed, “full parity” remains false.