VISUS Frozen Evidence bundle gate

A completed VISUS validation suite is not, by itself, sufficient for publication in GazeForge's Frozen Evidence layer. The lineage-integrity gate is v3 and requires the complete protocol-bound lineage introduced by the guarded VISUS workflow:

1. a pre-execution protocol frozen before detector inference;
2. the protocol-bound Grounding DINO + SAM 2 prediction batch;
3. the protocol-bound model-human validation binding;
4. the isolated source-authority transition seal;
5. the authority-bound validation-suite clone; and
6. strict five-input raw-execution provenance with the reviewed authority-certificate semantics revalidated.

A legacy suite plus execution-provenance pair is therefore no longer sufficient for scientific-review eligibility. Passing v3 is also not sufficient for public Frozen Evidence publication: the dashboard additionally requires a separate explicit scientific-review approval bound to the exact v3 lineage.

Public Python entry points

The v3 eligibility gate remains:

from gazeforge import (
    VisusFrozenEvidenceBundle,
    load_visus_frozen_evidence_bundle,
    validate_visus_frozen_evidence_bundle,
)

summary = validate_visus_frozen_evidence_bundle(
    "path/to/authority-transition"
)
bundle = load_visus_frozen_evidence_bundle(
    "path/to/authority-transition"
)

The authority-transition directory must contain:

• visus-dynamic-aoi-suite-manifest.json;
• every child report referenced by that suite;
• visus-protocol-authority-transition.json; and
• visus-execution-provenance.json produced by the strict authority-execution path.

The exact #130 visus-protocol-bound-validation.json remains in the original protocol-validation directory. The v3 gate reads the binding filename and exact file SHA-256 frozen by the #131 transition and searches the authority-transition directory's parent tree for one unique byte-identical binding. If the layout contains multiple identical copies, Python callers can remove the ambiguity explicitly:

summary = validate_visus_frozen_evidence_bundle(
    "path/to/authority-transition",
    protocol_validation_binding_path=(
        "path/to/validation/visus-protocol-bound-validation.json"
    ),
)

This lookup does not modify or relocate the #130 validation directory.

Strict command-line eligibility path

gazeforge-visus evidence-validate /path/to/authority-transition

evidence-validate deliberately has no suite-only, execution-only, or legacy-v2 eligibility mode. It requires the v3 lineage to resolve completely. Lower-level validation commands remain useful for diagnostics, but they are not substitutes for the review-eligibility gate. A successful evidence-validate result does not itself authorize dashboard publication.

What v3 verifies

The gate first validates the authority-bound suite and every referenced child report. It then validates the #131 transition seal and requires:

• the transition schema, status, and deterministic fingerprint to revalidate;
• all transition claim-boundary flags to remain fail-closed;
• the post-authority suite fingerprint, report count, and exact manifest bytes to match the current suite;
• the transition's pre-authority projection fingerprint to equal the frozen pre-authority suite fingerprint;
• source-audit/spec/manifest and authority-certificate identities to match the authority-bound suite;
• every child report in the transition ledger to have the same filename, report fingerprint, and exact byte SHA-256 as the current authority suite.

The gate then resolves and validates the #130 protocol-validation binding. It requires:

• schema gazeforge-visus-protocol-bound-validation-v1 and status verified-protocol-bound-validation;
• the binding fingerprint to recompute;
• the binding's exact file SHA-256 and semantic fingerprint to match the #131 transition;
• the pre-execution protocol fingerprint and protocol-batch fingerprint to match the transition;
• the binding's validation-suite fingerprint to identify the transition's exact pre-authority suite;
• source-audit/spec/manifest identities to remain unchanged across the authority transition;
• prediction_emission_grid_used=false in the frozen evaluation settings;
• model-human validation to have executed;
• source authority to remain a separate gate at the #130 layer; and
• all #130 claim-promotion fields to remain false.

Finally, strict authority-execution provenance is revalidated and must bind exactly five raw inputs, including the reviewed authority-certificate bytes. The execution suite fingerprint and authority-certificate fingerprint must equal the authority-bound suite identities, and certificate semantics must revalidate under the current closed schema.

A successful result reports:

bundle = "visus-frozen-evidence-v3"
status = "verified-protocol-authority-bound-bundle"
protocol_bound_lineage_verified = true
frozen_evidence_eligible_for_scientific_review = true
scientific_review_completed = false
empirical_performance_claim_created = false
formal_preregistration_verified = false

The typed VisusFrozenEvidenceBundle also carries the transition fingerprint, protocol-validation binding fingerprint, pre-execution protocol fingerprint, protocol-batch fingerprint, suite fingerprint, execution fingerprint, and source-authority certificate fingerprint.

Separate scientific-review approval

Public VISUS Frozen Evidence requires an additional file in the authority-transition directory:

visus-scientific-review.json

The review artifact is a separate closed-schema decision record. It is not generated from model metrics. A reviewer must explicitly supply reviewer identity, an ISO-8601 UTC review timestamp, and a non-empty review rationale after scientific review has occurred.

The Python review path is intentionally separate from v3 eligibility:

from gazeforge.visus_scientific_review import (
    build_visus_scientific_review_approval,
    validate_visus_scientific_review_approval,
    write_visus_scientific_review_approval,
)

record = build_visus_scientific_review_approval(
    "path/to/authority-transition",
    reviewer="Reviewer identity",
    reviewed_at="2026-09-11T18:30:00Z",
    review_rationale="Documented scientific review rationale.",
)

write_visus_scientific_review_approval(
    "path/to/authority-transition",
    reviewer="Reviewer identity",
    reviewed_at="2026-09-11T18:30:00Z",
    review_rationale="Documented scientific review rationale.",
)

validated = validate_visus_scientific_review_approval(
    "path/to/authority-transition"
)

The approval record binds the exact:

• authority-bound suite fingerprint;
• pre-authority suite fingerprint;
• strict execution fingerprint;
• protocol-authority transition fingerprint;
• protocol-validation binding fingerprint;
• pre-execution protocol fingerprint;
• protocol-bound prediction-batch fingerprint;
• source-authority certificate fingerprint;
• report count; and
• required five-input execution cardinality.

The record may promote only these publication-state fields:

scientific_review_completed = true
approved_for_public_frozen_evidence = true

It must keep all of the following false:

empirical_performance_claim_created
formal_preregistration_verified
independent_human_streams_verified
human_reference_ground_truth_promoted
source_authority_or_rights_expanded
raw_source_redistribution_authorized
evaluation_grid_boundary_changed
universal_performance_validity_claim_created

Changing any lineage identity invalidates the approval even if the review record is re-fingerprinted. Likewise, changing a prohibited claim to true remains invalid after re-fingerprinting. Existing review files are protected from replacement by default.

Dashboard behavior

The public benchmark dashboard now enforces three linked VISUS publication checks:

1. v3 lineage eligibility must validate completely;
2. visus-scientific-review.json must validate as an explicit approval for the exact same suite; and
3. every benchmark-shaped VISUS model-human or independent-human result row must be the exact named child recorded by that approved suite, with the same report fingerprint and the exact manifest-relative path.

A suite is therefore not surfaced merely because its completion manifest, child reports, transition seal, protocol-validation binding, authority certificate, and raw-execution provenance are internally valid. Those artifacts establish review eligibility. Publication requires the separate scientific-review decision.

The third check closes the detached-child route. A copied or separately surfaced VISUS result JSON cannot enter Frozen reports merely because its deterministic benchmark fingerprint is internally valid. The dashboard first classifies known VISUS child schemas using both evaluation_type and validation_scope; unknown or inconsistent VISUS- result schemas fail closed. It then validates the sibling reviewed suite and requires the report's child name, fingerprint, and exact relative path to match the approved suite inventory. Generic non-VISUS benchmark reports retain the existing integrity-only publication path.

An older VISUS directory containing only a suite manifest and visus-execution-provenance.json fails closed. A complete v3 directory without a scientific-review approval also fails closed instead of appearing under Frozen benchmark evidence. A detached VISUS child without the approved suite fails closed as well. Lund and other non-VISUS suite paths are unchanged.

What eligibility means

Eligibility is intentionally narrower than empirical validity. It means that the artifact chain is complete and internally consistent enough to be considered for scientific review. It does not mean that GazeForge has independently established that:

• a local VISUS copy is a current authoritative distribution;
• analysis or redistribution rights exist beyond the separately reviewed source-authority evidence;
• a human reference stream is ground truth;
• the historical two-contributor curation process represents two independent annotation streams;
• Grounding DINO + SAM 2 performance is scientifically adequate;
• model-emission frames can serve as the evaluation timestamp grid; or
• the pre-execution protocol constitutes formal preregistration.

Formal preregistration would require separate trusted temporal evidence. Independent human-stream evidence, empirical performance interpretation, source-authority/rights review, and final scientific review remain separate gates.

Even an approved public Frozen Evidence record is scoped only to GazeForge's reviewed evidence index. It does not create a universal performance-validity claim, convert the human reference into ground truth, expand source rights, authorize redistribution, or establish formal preregistration.

Why v2 and v3 eligibility alone are insufficient

The previous v2 gate required an authority-bound suite plus strict execution provenance. That was a strong raw-input and authority-integrity check, but it did not require the newer #130 protocol-validation binding or #131 transition seal. A caller could therefore reach review eligibility without proving that model inference and validation were tied back to the pre-execution protocol.

v3 closes that lineage bypass. #133 closes the subsequent suite-publication bypass. #135 closes the remaining detached-child publication route. The complete public path is now:

pre-execution protocol → protocol-bound prediction batch → protocol-bound model-human validation → isolated authority transition → authority-bound suite → strict five-input execution provenance → v3 scientific-review eligibility → explicit scientific-review approval → exact approved suite-child membership → public Frozen Evidence dashboard.

Every arrow before the final approval is an integrity/provenance statement. The approval is an explicit scoped review decision, and the child-membership check only proves that a displayed VISUS result is one of the artifacts that decision actually covered. Neither step automatically creates a scientific performance claim.