Verify a NoireBox export journal in CI. The job fails if the chain was tampered with.

The flight recorder meets the pipeline: any repository that consumes AI-agent decisions can gate its builds on the integrity of the journal it received — without trusting the producer, the host, or anyone.

The standalone NoireBox verifier recomputes the entire hash chain (SHA-256), checks every Ed25519 signature, validates the signed attestation and the RFC 3161 anchors — then exits 0 (INTACT) or 1 (TAMPERING DETECTED, with the exact event number).

- uses: noirebox/noirebox-verify@v1
  with:
    export-path: export.json   # the journal you received from a NoireBox instance
    noirebox-ref: main         # optional — pin a tag for reproducibility

The result output is INTACT or TAMPERING for downstream steps.

From any NoireBox instance (API, SDK or MCP):

$ curl -s https://your-noirebox.example.com/api/v1/export > export.json

The export contains every sealed event plus the signed attestation — it is the same dossier a third party would verify offline. One file, one command, zero trust.

• Light by design: the verifier needs only cryptography at runtime — no scikit-learn, no model download. The action clones the pinned NoireBox ref, installs that single dependency, and runs.
• Separable layers: this action proves integrity, not business completeness — the verifier's contract stays exactly what the threat model says it is.
• Same tool, both sides: the producer seals with NoireBox; the consumer verifies with the same verifier. Nobody imports a "trust me" SDK.

MIT — see LICENSE. Part of the NoireBox project.