Verifiable Model Record: Offline Provenance for AI Models

Field notes

The proposed record separates model provenance from the model itself: it identifies files, training input, lineage, production environment, and issuer through hashes and signed claims rather than exposing weights or training data. Verification is intended to work against locally chosen issuer keys, while policy packs would evaluate the record against named frameworks and return compliant, non-compliant, or indeterminate. The project says its version 0.1 specification and reference implementations are still pending.