GazeAudit publication and validation summary

GazeAudit is a methods-first Python package for auditing how eye-tracking conclusions change under measurement uncertainty and defensible analytical choices. The scientific MVP is validated by known-truth tests plus three deliberately different real-data outcomes. These outcomes are canonical, post-outcome frozen records; this document summarizes them but does not replace their protocols, source locks, execution records, or archived artifacts.

Scientific validation claim

The validation programme demonstrates that GazeAudit can represent three distinct scientific states without optimizing for a preferred result:

Case	Methodological stress test	Canonical outcome
GazeBase	seven-detector specification-space completeness	`incomplete`
Korthals target tracking	propagation of gaze measurement error into a paired AOI endpoint	`robust_negative`
Pedrotti/de Chambrier reading numerals	sampling-rate and added-missingness sensitivity of gaze-path rate	`materially_fragile`

The common integrity rule was protocol/source control before outcome inspection, checksummed execution artifacts, archive before reveal, and no post-outcome scientific tuning.

Canonical fingerprints

GazeBase multi-detector audit

protocol fingerprint: 3f64122f62cbc9762b0bd0e0b95c7fef6ff40c4700215ee4b90f005af77003b1;
scientific commit: 609203dfc37f2ef4984de82b888660e7cc30675c;
execution fingerprint: 35975d89dea728bd3aa146739dc58ea2bcb9322e357a44856b424fffb4183632;
artifact-manifest fingerprint: 965ba4b83b06bd3da549b9ca9e36e79443e33b2bbde76d81f87ed7767f94f71d;
archived artifact ID: 10284017265;
artifact ZIP SHA-256: 2fae4e99243e6738047a34b8dc24a183e8fb98606e4873723d43b8baa4ae937b;
canonical outcome: incomplete.

GazeBase does not have a separate source-freeze workflow. Its selected source-file byte identity is hashed and bound within the controlled execution/source identity. The failed completeness gate is retained as the scientific result; NH and REMoDNaV are not silently removed.

Korthals target-tracking AOI audit

protocol-v2 fingerprint: 1cd4150c9341db145b8f6aee544f9634c67c46116e61e7e675ae3c50282d2d55;
source-lock fingerprint: 8bcd9e1dfc8d6ca07c6222a20f563bf8b3bd67aed0de0d3b0e97628799de3a08;
scientific execution commit: 3312423cb43b1d9da3877914455dd471aba81174;
execution fingerprint: 2db2b92fea136ba516f38a87a3c218cf90c9a2a69c55e8ea37b63cad8febb51b;
audit scientific fingerprint: aa4c04af9c011fe997ef107c8250b20d6a20824b180abdeb16871b5b5a2fc768;
artifact-manifest fingerprint: c55dbeb476197196a36d1c1af857bd6424e2db6fb1c7c5e5bdc94169bd380de0;
source-freeze artifact ID: 10305414874;
scientific artifact ID: 10306308440;
scientific artifact ZIP SHA-256: 2eaab8171d9d5f70e138eb929554648cb6bed10b21646ab0acd7cfae4d607d11;
canonical outcome: robust_negative.

The 95% interval reported for this case is induced by the frozen measurement-error model. It is not a population confidence interval or Bayesian posterior interval.

Pedrotti/de Chambrier sampling + missingness audit

protocol fingerprint: efb194ad2c492768962320198d281c98574eb2dc22291279bfd176a15606a0d5;
source-lock fingerprint: d71949f035ecdf14b9e225b8ca441d291625767dc4a07fa4854882ff2dcae898;
source-manifest fingerprint: ca676426a01c1d57dda93b62709d4e9040369e4193c0b44353be78d644a33b1f;
scientific execution commit: 270d444c599363ebd25f25e8a95f5f3c78e55b02;
results fingerprint: 5bafe0d1d1d0d054a681cb7136fc4e259c0f75e07d51653acde472c7c6e4dce9;
execution fingerprint: 35dd53da10646400aca657c388f8b77a6a974c555de7efbd6f83faeb970b1eb6;
artifact-manifest fingerprint: 05a986abe8a7be7a7ccbf2bf24b7467a3c623381d05d2dab8f4ae81b0618d661;
source-freeze artifact ID: 10326523152;
scientific artifact ID: 10328414078;
scientific artifact ZIP SHA-256: b411fe6a5f4ad7035ac606d60a952306c0b818d956f7b4930827b8cbe10ae8d5;
canonical outcome: materially_fragile.

The reference effect remained directionally negative across all frozen sampling rates, but recovery failed the predeclared magnitude criterion at lower rates and in several added-missingness families. The classification therefore reflects the frozen recovery rule rather than a significance threshold.

Reproducibility map

The machine-readable evidence map is REPRODUCIBILITY_INDEX.json. The authoritative publication-facing matrix remains VALIDATION_MATRIX.md. Protocol, source-lock, execution, and result paths in the JSON index are repository paths and are regression-tested for existence and canonical identity.

Release-readiness evidence

This publication bundle was prepared from certified pre-release main c5f85d6f58ec058c29953b4513dbf3995f390f06, package 0.1.0.dev20. Exact-main tests #340, interoperability #255, distribution #2, and Pages #17 were green. The clean distribution build reproduced:

wheel SHA-256 debd6f6dcbdb337f15890d6124c73274e78db84ac36a79504b03297c4d56c9d0;
sdist SHA-256 3118d6b219be7ed549dc8689eb167cd9ecb3a96f30ce38b21acdad42fc310361;
distribution artifact ID 10341022917;
artifact archive digest sha256:c8ff5245ca6ce2a79c409257bf605acb4f7f1afa90c7705058c8cbe3ba3858f9.

These are pre-release qualification records, not a claim that 0.1.0 has been tagged or published.

Interpretation boundary

The real-data cases demonstrate behavior under their declared protocols and source identities. They do not establish universal robustness of every eye-tracking endpoint, tracker, population, preprocessing pipeline, or uncertainty model. Post-MVP additions must be validated separately and must not retroactively alter these canonical results.