license: apache-2.0
pretty_name: GSPC-PRV — Article 50 marking survival, validity principle (v3)
tags:
- ai-governance
- benchmark
- eu-ai-act
- provenance
- c2pa
- measurement
configs:
- config_name: default
data_files:
- split: train
path: items.jsonl
GSPC-PRV v3 — provenance survival by validity, not by presence
n = 16 public (+ 8 held out privately). v2 was measured DEAD — frontier models aced the C2PA-awareness matched pairs at 0.93–1.0. v3 targets the harder, correct principle.
What this measures
A marking SURVIVES only if a validator would report a VALID Content Credential bound to the asset; a manifest that is merely present but whose hard binding or signature no longer validates has NOT survived. Items pit present-but-invalid cases (a manifest copied verbatim into re-encoded pixels; pixels changed without a re-sign; a broken content hash; a tampered signature; an unresolvable external manifest) against valid-despite-transform cases (aware re-sign after crop/transcode/rotate; byte-faithful copy; valid lineage via ingredient). A model that pattern-matches "a manifest is attached, so it survived" fails the present-but-invalid half.
The grading instruction states this validity rule explicitly, so the labels are defensible (not a contested-label set). Whether v3 discriminates is an empirical question answered by the board — this is one cycle of author → measure → iterate, published honestly whichever way it lands.
Provenance
Authored by CSOAI (Apache-2.0, no third-party data), disjoint from published benchmarks by construction. Canary GUID in row 1; 8 held back privately. Contamination guard confirmed zero overlap with our training corpora before publication.
Grading
Deterministic SURVIVES/DESTROYED extraction. Unreadable → UNMEASURED, excluded, never scored wrong.
Honesty register
Measurement, not certification. n is below usable_n=30 until the bank grows; report the n with any figure. Nothing here is legal advice. CSOAI Ltd (GB, Companies House 16939677) · csoai.org