Verifiable Evaluation

Build provenance-oriented evaluation runs with attestations, signatures, and transparency evidence.

What this covers

  • Deterministic run artifacts (records.jsonl, manifest.json, summary.json, report.html)
  • DSSE attestation generation (insidellms attest)
  • Sigstore signing (insidellms sign)
  • Signature verification (insidellms verify-signatures)
  • Optional transparency and distribution paths (SCITT receipts, OCI publish)

Prerequisites

  • A completed run directory (must include manifest.json)
  • cosign installed for signing/verification workflows
  • Optional: oras for OCI publishing workflows

Use:

insidellms doctor --format text

to check common readiness items (ultimate:tuf, ultimate:cosign, ultimate:oras).

Quickstart: attest → sign → verify

# 1) Generate DSSE attestations for an existing run
insidellms attest ./baseline

# 2) Sign generated attestations
insidellms sign ./baseline

# 3) Verify signature bundles
insidellms verify-signatures ./baseline

To enforce signer identity constraints:

insidellms verify-signatures ./baseline --identity "issuer=https://token.actions.githubusercontent.com"

Expected run-directory additions

After attest:

  • attestations/*.dsse.json

After sign:

  • signing/*.sigstore.bundle.json
  1. Generate deterministic baseline/candidate runs.
  2. Generate attestations for both runs.
  3. Verify signatures in CI.
  4. Run insidellms diff --fail-on-changes as a behavioral gate.