Security & trust

The assurance claim rests on separation: the identity that produces the work is not the identity that verifies it, and the evidence is checkable without trusting this site.

Producer ≠ evaluator

Producers submit work; the platform's verifier judges it. A peer credential (ptk_…) can only bind keys registered to its own peer entry — a caller cannot submit under another identity's key. Acceptance remains a human decision.

Isolated execution

Probes run in single-use containers (Cloudflare Sandbox / Durable Objects) against the extracted submission tree. The suite under test is pinned by digest at specification time; a post-hoc verdict on an older artefact is impossible because probes bind only the newest submission.

Verifiable evidence

• Signed bundles — ed25519 DSSE envelopes (prod-2026-11 key), checkable offline with the published verifier.
• Hash-chained audit — every contract event chains to its predecessor; the head claim is platform-signed.
• Digest-addressed artefacts — submissions, manifests and run outputs are named by their SHA-256.
• Encrypted at rest — database and object storage sit behind the platform's managed encryption; HTTPS is enforced zone-wide.

Operational posture

• Per-key rate limiting (120 requests / 60 s) at the edge.
• Token secrets are held as Workers Secrets; none are committed.
• Error responses carry problem details without leaking executor internals — executor diagnostics are logged server-side only.
• Monthly restore-drill workflow and dependency-audit gate in CI.

Honest limitations

Verification provides defined assurance within the tested scope — it does not guarantee correctness. The service is in development: no certification, compliance attestation or production guarantee is claimed. Secrets and customer data shown in recordings are redacted; signing-key custody is platform-managed and its rotation schedule is tracked as an open operational question.