Cover security and operational gates

Separate runtime and operator roles for offline+online eval harness for grounded support answers. Runtime may only perform the narrow actions that ML engineer gating a prompt change before Friday release needs; operators get audited break-glass with TTL. Encode a negative probe that denies the privilege trick related to eval set leaked into few-shot examples — scores look perfect, prod drops.

Redact prompts/PII at collection. Secrets enter via a manager or workload identity — never source, fixtures, or exception strings. Incident EVAL-LEAK-308 should be impossible if these controls hold. Output allowlists and schema checks stay in force on error paths.

Rehearse the rollback or kill switch timed against a clock. Record actor, reason, prior revision/secret/flag, and verification query. Invariant reminder: release needs groundedness ≥ 0.88 and latency p95 ≤ 2.0s on fixed gold set v12.

Abuse cases unique to this lab include the privilege path implied by eval set leaked into few-shot examples — scores look perfect, prod drops. Prove a read-only role cannot mutate. Break-glass tokens expire; leftover tokens fail the drill. Dependency pin/digest story matters when images or models move under you. Document how to rotate the credential that offline+online eval harness for grounded support answers uses without a full outage window longer than your dual-run plan. Security evidence is part of ship, not an appendix nobody reads.

List network egress destinations and justify each. Ensure debug endpoints are off by default in the shipping config. Verify that error responses do not echo secrets or raw stack frames to clients. For multi-tenant paths, add a cross-tenant probe fixture. Time the rollback drill twice — once with the author, once with a peer. Store the drill transcript beside the threat notes for the incident id. In this chapter the product is offline+online eval harness for grounded support answers, the human stakeholder is ML engineer gating a prompt change before Friday release, and the incident id you design against is EVAL-LEAK-308. Re-state the oracle in your notes — candidate prompt beats baseline on groundedness by ≥ 0.03 without latency regression > 10% — and keep the invariant visible: release needs groundedness ≥ 0.88 and latency p95 ≤ 2.0s on fixed gold set v12. Track groundedness, citation_precision, p95_latency_ms as the scoreboard. Surface under change control: python -m eval.run --gold v12. If you only have forty minutes, finish the fixture for eval set leaked into few-shot examples — scores look perfect, prod drops before polishing UI. Promotion language stays ternary: promote, hold, or roll back based on evidence, not hope.