Brief
If you cannot name the failure modes, the acceptance cases, and who can pause the system, you are not ready for production — no matter how good the demo looks.
Write down the decisions the system will influence. For each, list acceptance examples and known failure modes. Pull them from production logs and operator interviews — not from marketing prompts.
If you cannot assemble that set, you are not evaluating a system. You are admiring a demo.
Every material change to prompts, models, retrieval, or tools should re-run the set. Store versions. Record who approved the release. When quality drops, you need a story that is not ‘we’ll look into it.’
For consequential writes — money, clinical status, identity, safety — require a reviewer and show them enough context to disagree. Automation without authority is how you earn incidents.
Monitor not only uptime but also score distributions, abstention rates, and user overrides. Name who can pause the system. Write the runbook before the first fire drill.