Skip to content

Gaze-in-the-Wild layout convergence

This evidence layer answers one narrow recovery question: what filename and MATLAB-variable layout is supported by the first-author processing code, and is that layout independently reproduced by downstream implementations?

It does not answer the stronger questions of whether a recovered tree is the original distribution, whether its bytes are authoritative, or whether the dataset files may be analysed or redistributed.

First-party layout

At pinned first-author repository commit 52262d44e366a53369e10ca73c5f41daf0e8f1e5, PlotLabels.m establishes the processing-code conventions:

PrIdx_%d_TrIdx_%d.mat
PrIdx_%d_TrIdx_%d_Lbr_%d.mat
ProcessData
LabelData

The reviewed PlotLabels.m Git blob is 511581250e04c62037c71d2da16271be4979d434.

The same first-author repository's pinned .gitignore blob, 85f9e994c92da6cbb5632ab241ae7473a83b35e5, excludes .mat and other data/media products. Combined with the already-frozen repository-history audit, this keeps the processing repository separate from the externally hosted ProcessData / LabelData distribution.

Accordingly, the verified filename and MATLAB-variable grammar is authoritative processing-schema evidence, not an authoritative dataset copy.

Independent downstream convergence

Three independent downstream repositories were reviewed at immutable commits.

Source Pinned revision Reviewed file Classification
DFKI OpenGazeLab 2c87abb3ed5d3e21ed027252a8fcd4fcfd9bdeee backend/src/preprocess_headmounted/giw.py Full filename + ProcessData / LabelData layout match
ACE-DNV 3142eb4457087743664d96994e952ed784741d1f modules/GiW.py Full filename + ProcessData / LabelData layout match
LEO-UMCG Unsupervised Gaze Event Discrimination 12ff306d8690e483d5e721e18d447fbd0d887e54 preprocessing.py Raw GIW label-file + LabelData corroboration only

The exact reviewed Git blobs are:

DFKI:     eebf405662caa0657095012d9451b3a0576bb5dc
ACE-DNV:  eb30815510c2e04d4239e06b7b22ffc74c10c595
LEO-UMCG: 7489c5df1ef5f5d7125f14b6b609ab2e170c0d3e

Two independent downstream implementations therefore reproduce the full first-party filename/variable layout, while a third independently corroborates reading raw GIW labeller files through LabelData.

This convergence increases confidence that the grammar is historically representative of GIW processing. It does not demonstrate that any downstream tree or archive is byte-identical to the original distribution.

Recovery-screening use only

A future recovered candidate may use these conventions as a non-authoritative screening signal. For example, a tree containing plausible PrIdx_*_TrIdx_*.mat and PrIdx_*_TrIdx_*_Lbr_*.mat paths can be flagged as layout-consistent for human review.

That flag must not become any of the following:

  • exact-copy identity;
  • first-party source authority;
  • original-distribution equivalence;
  • dataset-file analysis permission;
  • redistribution permission;
  • participant or task identity;
  • sampling-cadence verification;
  • independent-labeller recoverability; or
  • empirical-evidence eligibility.

The existing recovery quarantine therefore remains unchanged: file roles remain unclassified until the stronger evidence-gated transition is satisfied.

Frozen identities

The layout-convergence evidence record is bound to:

first-author repository-history evidence:
800d84d71d1d4b1a07e3b6d07c3bb7093c679284f49db0930a9836d77da30ad3

prior secondary-recovery-lead evidence:
e312079108f8b50ddedd6f361272218fc8665c880b147797aee5bb434ebc8c29

historical distribution-availability evidence:
2400c81a0897fb414285069c368a8a9d96de1d18eb185b1073cf15bb1c8bd1da

Frozen layout-convergence evidence fingerprint:

b9c006e7bc367d7a66a4e78577d4d267b488eec298619b7e2e6707468172ac12

Reviewed exact-file probe fingerprint:

98d9f5d81bed214248c08dbb3901b548b1ae6849f1fd85a33bec474c6e202943

The live workflow re-downloads the exact pinned source files, recomputes their Git blob identities, rechecks the required grammar, binds the probe to all frozen parent evidence, and reruns focused adversarial tests.

What remains unresolved

This tranche does not obtain an authoritative original or canonical GIW compressed distribution. Dataset-file rights, analysis use, and redistribution remain unresolved. Quarantine exit, source-audit readiness, participant/task mapping, actual distributed-file cadence, independent labeller-stream recovery, human-human agreement, participant-disjoint model validation, cross-dataset validation, GP3 validity, and Frozen Evidence performance claims all remain closed.

The next substantive milestone remains acquisition and independent review of an authoritative ProcessData / LabelData distribution together with applicable dataset-file reuse terms.