Design release criteria and trust review for an AI or ML product.

Instruction: Create a release decision record and staged operating plan from the supplied readiness evidence. Identify missing gates and who can authorize progress.

Context:

Fictional internal knowledge assistant. Approved pilot criteria: at least 90% task success on a representative held-out evaluation; no unresolved critical access-control defects; a verified human fallback; a tested disable switch; a named on-call owner. Current evidence: 920/1,000 tasks successful, no open critical access defects after a scoped security review, fallback documented but untested, disable switch last tested on the prior release, and no weekend on-call owner. Pilot exposure is capped at 50 consenting internal users. Product and security own the go/no-go; operations owns support readiness. The 90% rule is an observed benchmark criterion, not a population confidence claim.

Updated

Official answer available

Read the opening below, then unlock the full answer and practical guidance.

Observed task success is 92%, so the supplied benchmark gate passes. The scoped access-control review provides evidence for its gate but is not a guarantee of security. Operational readiness is incomplete: demonstrate the human fallback, retest the disable switch on this release and assign support coverage for the actual pilot schedule...

Upgrade to view official answer

Related Questions