Production integration checklist

Contract

• Pin API request/response version, model identity/version, preprocessing, and dataset/checkpoint identity.
• Test success, validation failure, timeout, unavailable model, fallback, and review routing.
• Preserve a client-compatible response shape and stable structured error codes.

Privacy and observability

• Log correlation ID, versions, route, latency, score band, review decision, and safe error code.
• Exclude media, frames, thumbnails, paths, recognized text, prompts, notes, and secrets.
• Test redaction, retention, access, and a public-safe log fixture.

Operation

• Record median/p95 latency, memory, hardware, runtime, and public fixture checksum.
• Label heuristic fallback as fallback; never convert failure into a negative prediction.
• Keep uncertainty and high-cost cases in human review. Prove the service degrades safely.

Release proof

• Run CPU contract tests and the local RTX 4080 graduation smoke separately.
• Walk the architecture and trust boundaries from approved input to reviewer.
• Save commands, versions, results, known limits, rollback point, and owner.