Skip to content

Gaze-in-the-Wild task-identity source recovery

This page records a bounded negative source-recovery result for the Gaze-in-the-Wild (GIW) benchmark. Its purpose is to answer a narrow question: does the pinned first-author repository contain an explicit task-to-TrIdx mapping that is strong enough to connect distributed PrIdx_*_TrIdx_*.mat files to the four publication tasks without inference?

Result

No explicit mapping was recovered from the reviewed first-party route.

The probe is pinned to:

  • repository: https://github.com/RSKothari/Gaze-in-Wild;
  • commit: 52262d44e366a53369e10ca73c5f41daf0e8f1e5;
  • 56 reachable commits;
  • 302 unique text blobs across the reachable history;
  • exact GIWApp.mlapp Git blob 2d560af336af637e042ea7a96f4c3736a59d2a2f.

The MATLAB app is a nine-member container. Its reviewed UTF-8 and printable-string surfaces contain zero task contexts. Across the reachable text history, only one blob is task-related: PlotLabels.m, which contains the local path fragment GIW_rearranged/Indoor_Walk. The recovered task contexts contain no nearby explicit TrIdx = N assignment. The deterministic probe therefore reports identity_plus_tridx_candidate_context_count = 0.

The frozen evidence fingerprint is:

caac3aebc99be005f8c2dcf431041e1d727b9ca07f2e4509c06d40b4f3452212

The discovery probe was produced by workflow run 34163709562, artifact 10033447214, at exact GazeForge head a7b21c151439e03ad29e4e157024f83062186bab.

What this closes

This tranche closes the reviewed pinned first-party plain-text / printable-string task-identity recovery route. Re-running the dedicated workflow checks the same exact upstream revision and requires the source inventory, MLAPP identity, task-context counts, and fail-closed scientific boundary to remain consistent with the frozen evidence.

This is useful negative evidence: it prevents future work from repeatedly treating the MATLAB app or reachable source history as though they contained an undiscovered explicit lookup.

What it does not establish

The result is deliberately narrower than “no mapping exists.” It does not exclude mappings that may exist in:

  • private or author-held materials;
  • deleted or unreachable Git objects;
  • external project storage or archived web material;
  • semantically encoded binary structures not exposed as reviewed printable strings;
  • the actual distributed data contents after exact-byte acquisition.

The Indoor_Walk path is therefore not promoted into a universal TrIdx = 1 rule. Likewise, publication label-status patterns and the Figshare LabelData filename inventory can constrain some participant/trial cells, but zero-labelled trials remain ambiguous between publication states such as not labelled and discarded. Those constraints are not a substitute for an authoritative file-to-task ledger.

Scientific boundary

This tranche creates no new empirical performance evidence. The following remain false:

  • universal TrIdx → task mapping verified;
  • complete distributed-file → task mapping verified;
  • participant-disjoint GIW model validation created;
  • GIW human-human agreement created;
  • cross-dataset validation created;
  • native GP3 validity established;
  • GIW quarantine exit authorized.

The next scientifically useful GIW work must therefore obtain stronger external evidence—most importantly exact distributed bytes and/or an authoritative task/file mapping—before participant-generalisation or cross-dataset performance claims can be created.

Reproduce

The dedicated workflow runs:

python scripts/probe_gaze_in_wild_task_identity_recovery.py \
  _giw_source \
  --output gaze-in-wild-task-identity-recovery-probe-v1.json

It then validates the live observation against the frozen gaze-in-wild-task-identity-source-recovery-evidence-v1.json and runs adversarial tests that reject attempts to promote the negative result into a universal mapping or model-validity claim.