Skip to content

Gaze-in-the-Wild ProcessData preflight

GazeForge now has a fail-closed structural preflight for a quarantined Gaze-in-the-Wild ProcessData MATLAB file. The preflight exists to answer one narrow question before a candidate is allowed anywhere near the empirical source-audit path:

Does this exact ProcessData object contain the structural fields consumed by GazeForge's ProcessData-side adapter, with internally compatible shapes and timestamps?

It does not certify the Gaze-in-the-Wild corpus, dataset rights, annotation recovery, or empirical usability.

First-party sample used for the frozen check

The frozen record is derived from the pinned first-author repository RSKothari/Gaze-in-Wild at commit 52262d44e366a53369e10ca73c5f41daf0e8f1e5. PR #78 established that the repository's sole reachable tracked archive, DataExtraction/all_preprocessing_steps.zip, contains one data-bearing member named exports/ProcessData.mat.

The preflight rechecks that exact object rather than trusting the earlier observation:

Property Frozen observation
Archive SHA-256 5deef95a4d847b7b21a37c5d746212549b63217bd64159bec72992d82b575956
Embedded member SHA-256 d633bbf0a3a9224b71e286ada10a63abe7761264eec7e8c25c94fcf0bbbafc63
Embedded member bytes 40,528,499
PrIdx 2
TrIdx 2
Stored SR 300 Hz
Timestamp count 106,225
Timestamp span 0.0–354.0851999999941 s
Median-difference inferred processed rate 299.9955942810773 Hz
ETG.POR shape 106,225 × 2
ETG.Confidence shape 106,225
ETG.SceneResolution 1920 × 1080
ETG.Labels present yes, 106,225 values
Top-level LabelData present no

The frozen preflight record fingerprint is d2ec4cc066302646e199b88d29e6f8c165b1ce735c61f1652b8e00b1511a75dd.

What the preflight verifies

preflight_gaze_in_wild_processdata() checks the exact file hash and byte size when expected values are supplied, then verifies:

  • a top-level ProcessData MATLAB variable;
  • positive scalar PrIdx, TrIdx, and SR values;
  • finite, strictly increasing ProcessData.T values and a timestamp-derived processed rate;
  • numeric ETG.POR with either N × 2 or 2 × N shape matching T;
  • ETG.Confidence length matching T, while preserving the adapter's ability to treat non-finite confidence values as invalid samples rather than structural corruption;
  • positive integer ETG.SceneResolution width and height;
  • optional ETG.Labels, which must match T when present.

The workflow reproduces these observations from the exact pinned first-author Git revision and requires the live record to equal the committed frozen record exactly.

Scientific boundary

The only positive scientific-governance conclusions from this tranche are:

  1. this exact first-party embedded ProcessData object passes the structural preflight; and
  2. the ProcessData fields consumed by GazeForge's coordinate adapter are structurally compatible for this one sample.

Everything stronger remains closed. In particular:

  • exports/ProcessData.mat is not claimed byte-equivalent to the historical distributed PrIdx_2_TrIdx_2.mat;
  • the full/canonical Gaze-in-the-Wild distribution has not been recovered;
  • ETG.Labels is not treated as separately distributed LabelData and does not recover independent human-labeller streams;
  • the stored 300 Hz value and the approximately 299.996 Hz timestamp grid describe this processed object only; neither establishes the published acquisition-hardware cadence or a corpus-wide sampling-rate distribution;
  • the presence and shape of ETG.POR and ETG.SceneResolution do not independently certify the coordinate semantics of the historical distribution;
  • participant/task mapping is incomplete;
  • dataset-file rights, analysis permission, and redistribution permission remain unresolved;
  • quarantine exit and full source-audit readiness remain false; and
  • no human-agreement, model-validation, cross-dataset, GP3, or Frozen Evidence performance claim is created.

Relationship to the full source audit

This preflight is intentionally weaker than GazeInWildSourceAuditSpec in gazeforge.gaze_in_wild_audit. The full empirical audit still requires authoritative LabelData and ProcessData manifests, exact file identities, verified reuse terms, explicit analysis-use permission, participant/task identity mapping, coordinate evidence, and the remaining source-audit invariants.

A future recovered corpus can therefore pass through this preflight while still remaining quarantined. Structural compatibility is useful evidence, but it is not authorization and it is not empirical certification.