Skip to content

Gaze-in-the-Wild roadmap evidence synchronization

This checkpoint synchronizes roadmap entries with evidence that is already frozen in the repository. It does not rerun the underlying science and does not widen any Gaze-in-the-Wild claim.

Version 2 extended the earlier rate/HHA synchronization with a narrowly scoped completion for the official public Figshare ProcessData + LabelData copy and its current deposit reuse terms. Version 3 separates the already-reviewed task-agnostic participant-disjoint derived-60-Hz validation and event-class sensitivity from the still-blocked naturalistic-task sensitivity component. The v1 and v2 records remain immutable historical provenance.

Scoped items now supported

Roadmap item Evidence-backed status Scope boundary
Per-file sampling-rate distribution Satisfied, with corrected wording 68 distributed ProcessData files; rate inferred from each file's strictly increasing processed T timestamp grid. This is a processed-stream timestamp-grid rate, not acquisition-hardware cadence.
Labeller-to-labeller sample/event agreement Satisfied, with corrected wording Six labeller pairs across the five distributed recordings that actually contain multi-labeller overlap. This is overlap-subset HHA, not full-distribution HHA.
Official public GIW copy + current reuse terms Satisfied, narrowly The Rakshit Kothari-authored Figshare ProcessData and LabelData original-publication deposits are identified, their CC BY 4.0 deposit terms are verified, and all 118 files were byte-verified against the frozen manifests. This does not establish byte-for-byte equivalence to the historical RIT-hosted archive.
Participant-disjoint task-agnostic model validation Satisfied 12 participants, 18 recordings, five participant-disjoint folds, 157,850 matched OOF analysis rows per model on a derived 60-Hz grid. No task mapping was used.
Event-class sensitivity Satisfied Frozen sample/event class results are available for blink, fixation, pursuit, saccade, and VOR. Pursuit remains an explicit failure case.
Naturalistic-task sensitivity Open task_fold_metrics and task_summary each contain zero rows. Task sensitivity remains blocked until authoritative distributed file → publication-task mapping is verified.

Frozen synchronization identities

The original v1 synchronization remains frozen at:

validation/evidence/gaze-in-wild/gaze-in-wild-roadmap-evidence-sync-v1.json

with fingerprint:

c0dc3182f7fa89d325e0be2406d6d1aaf6541ebb2e90c21f49ade23398909571

The v2 synchronization is frozen at:

validation/evidence/gaze-in-wild/gaze-in-wild-roadmap-evidence-sync-v2.json

with fingerprint:

8baf35fc97e59fa691a32e3e3d190e431bd1214a851d38f1877dce6d448fa215

The v3 synchronization is frozen at:

validation/evidence/gaze-in-wild/gaze-in-wild-roadmap-evidence-sync-v3.json

with fingerprint:

426ad10551a52a4b1cae9df9ae422336b591188466051235b5b38c8e808f7aa2

The v3 validator revalidates v2 and the reviewed exact participant-disjoint model evidence rather than replacing either scientific meaning.

ProcessData processed timestamp-grid rate ledger

The upstream rate ledger is:

validation/evidence/gaze-in-wild/gaze-in-wild-processdata-processed-rate-ledger-v1.json

Fingerprint:

1bd6643bd0c75faec9d070f74fb719f9178caa2da7e90bcbe356a0dbc5f38905

It contains 68 exact distributed ProcessData file identities. For every file it records the stored ProcessData.SR processing-rate field and the rate inferred as the median inverse delta of the strictly increasing ProcessData.T processed timestamp grid. The stored field is 300 Hz and the timestamp-inferred processed rates are correspondingly near 300 Hz.

This evidence is deliberately narrower than an acquisition-cadence statement. It does not verify the hardware acquisition rate, native 60 Hz evidence, or Gazepoint GP3 validity.

Distributed-overlap human-human agreement

The upstream HHA evidence is:

validation/evidence/gaze-in-wild/gaze-in-wild-distributed-overlap-human-human-agreement-evidence-v1.json

Fingerprint:

7e6d6180b01417d7be2e07483cc0c7e0feb86a1d08be6b8aeba03f48fdb27f02

It freezes sample-level and event-level agreement for the five recordings with distributed multi-labeller overlap, covering six labeller pairs. This is an overlap-subset result, not a full-distribution HHA claim.

Official Figshare copy and rights chain

The distribution-rights evidence is:

validation/evidence/gaze-in-wild/gaze-in-wild-figshare-distribution-rights-evidence-v1.json

Fingerprint:

25d4b56e1dc4888b034b5de91cb14f712aaa0e3470cde7a1246f30c3880e63c7

It binds Figshare project 74580 and the named Rakshit Kothari-authored deposits directly to their dataset-record licence, CC BY 4.0. The relevant original-publication deposits are:

  • ProcessData: DOI 10.6084/m9.figshare.11673645.v1;
  • LabelData: DOI 10.6084/m9.figshare.11673696.v1.

The exact-byte evidence is:

validation/evidence/gaze-in-wild/gaze-in-wild-figshare-exact-bytes-evidence-v1.json

Fingerprint:

dbf277d698266fe53e337a15fc342e9af1835460d2bcc69c3992635de8fa0a00

It verifies 118 files / 2,413,299,242 bytes across ProcessData and LabelData, with transport-independent stable exact-byte identity:

8d36884daf76e26ded3d31c32bd334bf8f358b49efdb7accbb20c6c569abdda0

Every downloaded MAT file was deleted after verification. Raw dataset bytes were not retained. ProcessData_cleaned was not downloaded and remains excluded because its own deposit description identifies it as outside the original publication data.

This is sufficient to close the official public Figshare copy/current deposit-rights roadmap item. It is deliberately not restated as proof that the current Figshare bytes are byte-for-byte identical to the historical RIT-hosted archive.

Participant-disjoint task-agnostic model evidence

The reviewed participant-disjoint evidence is:

validation/evidence/gaze-in-wild/gaze-in-wild-exact-participant-disjoint-model-validation-evidence-v1.json

Fingerprint:

b2fe85ec7e5d5cd425c0cd2593742bab835686c3f560d8a6c06e9c6d67dc547a

Its frozen scope is:

  • 12 participants, 18 recordings, and 5 participant-disjoint folds;
  • 157,850 retained analysis rows per model with identical OOF rows;
  • a derived 60-Hz analysis grid, not native 60 Hz or GP3 validation;
  • task_mapping_used = false;
  • task_stratified_validation_created = false;
  • task_fold_metrics = 0 rows;
  • task_summary = 0 rows;
  • event-class sensitivity is present.

The cross-run reproducibility signature is:

f8c8d27ddbb1fe065a15df18d53c9fa84544315ef57cf2783554b236839ff286

Floating benchmark outputs are compared under the frozen eight-decimal canonicalization policy; source/split/count/class/convergence/scientific-boundary identities remain exact.

Pursuit remains a required visible failure case

The v3 validator does not permit roadmap synchronization to hide the weak pursuit result. Frozen pursuit support is 5,696 samples and 327 reference events.

  • I-VT: sample F1 0; event F1 0; no predicted pursuit events.
  • RandomForest: sample F1 0.0768712070; event F1 0.0085836910.
  • ContextMLP: sample F1 0.0547112462; event F1 0.0029585799.

Accordingly, aggregate learned-model performance must not be presented as uniformly strong event recognition.

Roadmap wording frozen by the validators

The accepted scoped statements are:

Freeze the 68-file Gaze-in-the-Wild ProcessData processed timestamp-grid rate distribution; this is not acquisition-hardware cadence.

Freeze Gaze-in-the-Wild labeller-to-labeller sample-level and event-level agreement on the distributed multi-labeller overlap subset (five recordings); no full-distribution HHA claim.

Obtain/audit the official Gaze-in-the-Wild Figshare ProcessData + LabelData original-publication bytes and current deposit reuse terms (CC BY 4.0); raw MAT not retained.

Run participant-disjoint Gaze-in-the-Wild task-agnostic derived-60-Hz model validation and report event-class sensitivity.

The still-open companion statement is:

Report Gaze-in-the-Wild naturalistic-task sensitivity only after authoritative distributed ProcessData-file → publication-task mapping is verified.

A broader replacement—such as claiming task sensitivity, historical RIT archive equivalence, quarantine exit, native 60-Hz validity, or GP3 validity—fails validation.

Still open

This synchronization does not satisfy or authorize:

  • byte-for-byte equivalence to the historical RIT-hosted archive;
  • authoritative complete numeric TrIdx → publication-task mapping;
  • TrIdx 4 → Tea_Making by elimination;
  • task-stratified human-human agreement or model validation;
  • naturalistic-task sensitivity claims;
  • acquisition-hardware cadence verification;
  • native 60 Hz or Gazepoint GP3 validity;
  • cross-dataset validation;
  • quarantine exit;
  • raw MAT retention; or
  • new empirical model-performance claims.

The task-agnostic participant-disjoint derived-60-Hz benchmark and event-class sensitivity are complete frozen evidence. The naturalistic-task sensitivity roadmap component remains open until the authoritative file-to-task mapping gate is resolved.