Korthals protocol v2: zero-finite matched-cell amendment

Trigger

The endpoint-blind korthals-source-freeze #6 run on certified main failed before any AOI endpoint, Monte Carlo draw, effect estimate, or classification was executed. The source-readiness guard found one retained task trial with zero finite gaze samples after the already-frozen 50-Hz scheduled-grid selection: participant 88878fe6, trial 27.

Protocol v1 is retained unchanged with fingerprint 2b8e8da84182316d08ab173ad7d71a2ff09fc45bd84f875ae548832a288ba291. Its rule that every retained trial must have at least one finite scheduled sample therefore remains a valid record of the failed v1 source-readiness attempt.

External evidence check

Before changing any code, the published data descriptor and the exact companion implementation were checked. Korthals, Visser, and Kucharský (2026; DOI 10.1038/s41597-026-06963-4) define missing data as gaze samples with NA values and report 2.29% missing data for participant 88878fe6 across trials 1–144. Their technical-validation summary states that data quality is generally good except for participant 21db28aa, trials 82–144, which they explicitly recommend removing after an unrecalibrated break. The paper also documents that EyeLink blink samples and the 50 ms before/after each blink are set to NA by the default preprocessing.

The frozen companion commit remains 1d7ebec23e3fe20f6db952eefda1a0dd58ae09da. Its OriginalPreprocessor implements that ±50 ms blink masking and does not provide a special exclusion or repair for participant 88878fe6 trial 27. No later upstream companion commit exists at the time of this amendment.

These checks indicate that the source-freeze failure is not an upstream instruction to change blink handling, grid phase, or participant eligibility. It is a conflict between ordinary published missingness and GazeAudit protocol v1’s stricter zero-finite-trial readiness rule.

Version-2 rule

Protocol v2 fingerprint: 1cd4150c9341db145b8f6aee544f9634c67c46116e61e7e675ae3c50282d2d55.

The only scientific amendment is general and symmetric:

Preserve the exact companion preprocessing, ±50 ms blink mask, target-master alignment, fixed 50-Hz grid, earlier-sample tie rule, validation mapping, AOI, matched-cell definition, participant/study weighting, Monte Carlo settings, and seed.
After fixed-grid selection, identify task trials with zero finite selected gaze samples.
Do not interpolate, substitute a nearby finite sample, move/phase-shift the grid, shorten the blink window, or special-case a participant/trial.
Instead, mark the affected paired cell structurally unestimable and remove both its moving_circle and jumping_circle members.
Preserve all unrelated matched cells.
Every source participant must retain at least one complete matched cell; otherwise preparation fails closed.
Validation mapping remains fail-closed before any missingness-based cell removal, so a zero-finite trial cannot hide an unmapped validation block.
Record the zero-finite trial identities and removed matched-cell identities in the endpoint-blind source-intake artifact.

This is a paired complete-cell rule, not an outcome-dependent exclusion. It is applied before any scientific effect is computed and is encoded for all participants/trials rather than for the observed 88878fe6/27 case alone.

Scientific boundary

At amendment time, no Korthals target-window occupancy effect, Monte Carlo endpoint distribution, sign-stability probability, or final classification had been executed or inspected. The source-freeze workflow remains archive-before-reveal and endpoint-blind.