Skip to content

Validation & evidence guide

GazeForge exposes several validation surfaces because implementation, empirical evidence, source identity, and reproducibility answer different questions. This page is a routing layer; it does not create a new evidence class or replace the generated evidence records.

Canonical evidence status

Evidence status is the canonical public status layer. If wording on this guide ever conflicts with that generated page, use the generated evidence status and its bound provenance.

Start from your question

  • What empirical evidence exists now?

    Start with the generated evidence classes and benchmark-specific boundaries.

    Evidence status →

  • Is a capability implemented?

    Check software/validation readiness without treating implementation as empirical validity.

    Validation status →

  • Which reports are frozen?

    Inspect publication-gated empirical reports, fingerprints, and provenance summaries.

    Frozen empirical evidence →

  • Is native 60 Hz / GP3 validated?

    Keep native-device evidence separate from lower-rate conditions derived from higher-rate corpora.

    Native 60 Hz validation → · Empirical execution →

  • Where did the evidence come from?

    Follow source-resolution state, authority, rights, and exact-byte provenance before making stronger claims.

    Source-resolution status →

  • What scope was actually certified?

    Use scope certificates to preserve dataset, split, rate, annotation, and execution boundaries.

    Validation scope certificates →

  • Which benchmark should I inspect?

    Enter the benchmark documentation by research question rather than by internal file name.

    Benchmark guide →

  • How should I report or cite it?

    Record exact software identity separately from the empirical evidence that supports a scientific claim.

    Reproducible reporting → · Citation & attribution →

What can I claim?

Research question Start here Do not infer
What is the current evidence class for a benchmark? Evidence status implementation status ≠ empirical evidence
Is a method or workflow implemented and tested? Validation status implemented ≠ externally validated
Is a report publication-gated and fingerprinted? Frozen empirical evidence reviewed/bounded evidence ≠ Frozen evidence
Does evidence establish native 60 Hz / GP3 validity? Native 60 Hz validation derived 60 Hz ≠ native-device validation
Is the underlying source identity/authority resolved? Source-resolution status recovery candidate ≠ authoritative source
Can another researcher reconstruct the software/evidence state? Citation & attribution a DOI or package version alone ≠ validation strength

Boundaries that must survive navigation

Routing should make evidence easier to find, not stronger than it is.

  • Lund2013: lower-rate evidence is derived from native 500 Hz expert-labelled data; it is not native GP3/60 Hz validation.
  • Hollywood2EM: source-token-disjoint evidence is not participant-disjoint.
  • Gaze-in-the-Wild: current participant-disjoint evidence is task-agnostic; the complete authoritative numeric task mapping remains unresolved.
  • VISUS: current evidence is bounded partial public-derivative evidence; it does not establish full VISUS model validation, Frozen Evidence, or native GP3 validity.
  • Synthetic/demo outputs: examples and known-truth tests are not empirical benchmark evidence.

For the current machine-readable/public classification of those states, return to Evidence status.

If the answer is "not yet"

Open evidence gaps are documented rather than hidden:

This guide intentionally avoids copying headline metric values. Use Results gallery for readable figures and Frozen empirical evidence for the validated report-level record.