Mission 7 · Operations and recovery
Make the system supportable after the demo.
Connect correlation IDs across the whole flow, expose health checks, bound retries, add a review lane, and prove one replay without duplicate side effects.
Build sequence
- Carry one correlation ID across CRM event, meeting, stream, recording, transcript, score, and write-back.
- Expose health and readiness separately and track queue age, replay, provider latency, artifact access, transcript quality, evaluator variance, and cost without content leakage.
- Use a durable outbox, bounded retries, delivery ledger, and dead-letter or human-review lane for every external write.
- Implement manual replay that reuses the original idempotency key and reconciles provider and CRM results.
- Prove retention expiry or consent-withdrawal deletion, then verify inaccessible artifacts and redacted logs.
- Run one backup and restore or state-reconstruction drill and write a five-failure owner runbook.
Required proof
- Failure matrix, privacy-safe health output, cost signal, and correlation trace.
- Outbox and delivery reconciliation, deletion proof, recovery runbook, and restore evidence.
- One successful replay with no duplicate meeting, artifact, delivery, or CRM write-back.
Reference files and sources
Official vendor documentation
Assessment gate
Pass
Pass: the operator can detect, explain, reconcile, delete, restore, replay, and recover the flow without duplicate side effects or private-content telemetry.
Revise
Revise: monitoring exists but outbox delivery, replay, deletion, restore, reconciliation, ownership, or failure guidance is incomplete.
Hard fail
Hard fail: unbounded retries, destructive replay, retained withdrawn media, private-content analytics, silent failure, or unverifiable recovery is present.
Human-reviewed submission