Skip to content

Hollywood2 participant-crosswalk exhaustion

GazeForge now freezes the result of a public-source search for an explicit mapping between the Hollywood2EM three-digit filename tokens and the original Hollywood-2 participant identifiers/task groups.

Result

No authoritative public crosswalk was recovered.

That result does not mean the tokens are unrelated to participants. It means the available evidence is insufficient to claim what each token identifies. The values therefore remain opaque source tokens.

Public surfaces reviewed

The exhaustion record binds the following evidence layers:

  • the original Hollywood-2 eye-tracking distribution page, which reports 16 volunteers split into 12 active action-recognition and 4 free-viewing subjects and refers to an included README, but does not expose a public subject ledger or GIN-token mapping on the description page;
  • the original academic-use licence/access surface, without bypassing registration, login, licence acceptance, or institutional access controls;
  • arXiv:1312.7570v1 and the final IEEE TPAMI article (10.1109/TPAMI.2014.2366154, PubMed 26352449), whose 16-versus-19 participant-cardinality difference is already frozen separately and cannot resolve token identity;
  • the 2020 Hollywood2EM paper (10.16910/jemr.13.4.5), which identifies 56 annotated clips viewed by 16 observers and links the canonical GIN data set, but does not provide an explicit filename-token → original-subject/task-group crosswalk;
  • the already complete seven-commit canonical GIN-history audit, including all three README versions, which recovered no participant/identity mapping from repository history;
  • a secondary public Hollywood2EM parser used for eye-event evaluation, which preserves Hollywood2EM paths but does not interpret the filename prefix as a participant ID or task group and is not authoritative for original participant identity.

Numerical coincidences remain non-evidence

The canonical GIN token inventory is:

001, 002, 003, 004, 005, 006, 008, 010, 011, 012, 013, 014, 015, 017, 018, 019

The missing integers inside 001–019 are 007, 009, and 016.

Two tempting shortcuts are prohibited:

  1. 16 GIN tokens = 16 earlier/original participants. Matching cardinality does not prove that each token is an original participant identifier.
  2. 3 missing integers = the 16→19 publication delta. The missing values are not evidence for the identities of three additional participants, nor for a task group.

Likewise, ordinal labels such as free-viewing subjects 1–4 in publication analyses are not equated with GIN tokens.

Frozen record

The reviewed exhaustion evidence is:

validation/evidence/hollywood2/hollywood2-participant-crosswalk-exhaustion-evidence-v1.json

Evidence fingerprint:

5900a3515243f92ea374a0dc5d1d80e6c95d64b1dc6fa7b4f698df85488eaafa

The corresponding validator is fail-closed. Refingerprinting an edited record cannot promote participant identity, task-group identity, participant-disjoint validation, cross-dataset validation, Frozen Evidence, native-60-Hz/GP3 validity, rights scope, or raw-data redistribution.

Remaining authoritative routes

There are now only three legitimate resolution routes:

  1. Obtain gaze_hollywood2.zip through the original distribution's normal institutional terms and inspect only its README/metadata with the existing metadata-only participant-ledger intake. An original subject ledger alone still does not prove a GIN-token crosswalk.
  2. Recover a public author/institutional source that explicitly maps every relevant Hollywood2EM token to an original subject identifier and task group.
  3. Obtain an explicit author-supplied token ledger through a normal authorized channel and independently review it before promotion.

No login bypass, raw-data redistribution, count inference, filename-shape inference, or third-party reconstruction is an acceptable substitute.

Current scientific boundary

Hollywood2EM validation remains source-token-held-out, not participant-held-out. Participant-generalization claims and Lund↔Hollywood2 cross-dataset validation that depend on verified participant semantics remain blocked until an explicit authoritative crosswalk is recovered and reviewed.