Production Readiness Review

Reliability

• Are customer outcomes and safety represented in the SLOs?
• Are deadlines and retry policies explicit for every dependency?
• Are unsafe side effects idempotent?
• Can partial workflows resume from durable state?

Operability

• Can one customer request be reconstructed from a trace ID?
• Can failures be sliced by customer, SDK version, runtime, endpoint, region, and deployment?
• Is there a standard evidence packet and escalation threshold for issues that cannot be reproduced locally?
• Do alerts identify customer impact and an owner?
• Are secrets and customer data excluded from logs?
• Is there a tested incident runbook?

AI quality

• Are outputs and tool arguments validated?
• Do evals cover regressions and critical slices?
• Are prompt, model, tool, and dataset versions recorded?
• Is there a safe fallback or human escalation path?

Data and side effects

• Is the source of truth explicit for each workflow?
• Are payment, webhook, and other side effects idempotent?
• Can partial success be reconciled without repeating completed work?

Release

• Are canary guardrails written before rollout?
• Can the team disable the behavior and roll back quickly?
• Has recovery been rehearsed rather than assumed?

Final decision

Choose ready, ready with conditions, or not ready. List evidence for the decision and name the owner and deadline for every condition. Production readiness is an accountable claim, not a feeling.