Model release checklist

Run on every model version change. Date and sign each line.

• New model ID pinned in config; old ID retained for rollback.
• Incumbent model run twice on the full suite; noise band recorded: .
• Capability suite run on the new model with its own best prompt; per-category table attached.
• Regression suite run; cases tagged regression vs capability.
• Safety suites run: direct injection, indirect injection, tool abuse, exfiltration, abstention. Per-category results attached.
• Any category down by more than 2x noise: STOP and record why, or fix and re-run.
• Cost per request and p95 measured on the new model.
• Canary on % of traffic for ; dashboard watched.
• Regression layer re-baselined after full migration.
• One paragraph on what changed in the model's behaviour, added to the yearly plan.