Model Provenance

Every model artifact is bound to a verifiable digital identity using short-lived certificates issued via OIDC. This ensures that the author, training pipeline, or signing service is authenticated before an artifact is published. Sigstore and Cosign enable keyless signing, eliminating long-lived key management risks. The binding creates a non-repudiable attestation that a specific, trusted actor produced the model at a specific point in time.

Provenance is captured as a set of cryptographically signed in-toto attestations stored in an append-only transparency log. These attestations form a verifiable, end-to-end record of every step in the model lifecycle:

• Source attestation: Links the model to a specific commit hash and repository.
• Build attestation: Records the training environment, dependencies, and hyperparameters.
• Evaluation attestation: Captures benchmark results and safety evaluations. Any tampering is immediately detectable through log consistency proofs.

A model's provenance includes a complete, queryable dependency graph of all upstream artifacts. This extends beyond code libraries to include:

• Base model weights and their own provenance records.
• Training datasets identified by cryptographic hash.
• Fine-tuning adapters and their version histories. This transitive visibility enables precise impact analysis when a vulnerability is discovered in any upstream component, allowing teams to instantly identify every affected model.

Provenance metadata captures the complete bill of materials for model creation, including pinned dependency versions, environment specifications, and training scripts. This enables a reproducible build process where an independent verifier can re-execute the training pipeline and confirm a bit-for-bit identical model artifact. Reproducibility transforms provenance from a claim into a provable property, eliminating the risk of tampering during the build process itself.

Provenance data feeds into automated policy-as-code engines like OPA that enforce deploy-time rules. Before a model can be served, the system verifies:

• The complete attestation chain is valid and unbroken.
• All signatures come from trusted identities.
• No known vulnerabilities exist in the dependency graph.
• The model has passed required evaluation gates. This binary authorization gate prevents unverified or tampered models from ever reaching production.

Model provenance is expressed using machine-readable standards to ensure interoperability across tools and organizations:

• CycloneDX and SPDX for software and data bill of materials.
• in-toto for step-level supply chain attestations.
• SLSA for graded integrity levels. These standards allow security scanners, policy engines, and artifact registries to consume and validate provenance data without proprietary integrations, creating an open ecosystem of verifiable AI supply chain security.