# Routing Regret

Routing Regret is a weekly candidate count, not an automatic verdict. A missing required signal returns `null`; zero means the data was sufficient and no task matched.

Conservative policy v1:

- Over-spawn: at least 2 delegated workers, verified success, known single-unit scope, complete attribution, zero-impact rate at least 75%, duplicate-finding rate at least 50%, no worker-supported decision, and coordination tokens at least equal to worker output tokens.
- Under-spawn: an initial DIRECT/SCOUT change or fix reaches green only after at least 12 root tool calls, 50% root input growth, 2 failed checks, 2 repairs, and more affected directories than initially identified.
- Model overkill: low-risk verified work with at most 3 root tool calls, no repair/escalation or attributed high-level decision, and at least 8 comparable tasks forming at least 5 cross-model pairs. A cheaper model must save at least 20% with the same verified and rework outcome.
- Model underpower: same-fingerprint escalation after at least 2 failed checks and 2 repairs, green within 5 minutes, complete credit telemetry, and at least 20% of the observed total cost spent before escalation.

Review candidates against the sanitized event log before proposing a versioned experiment. Never promote a config from this metric alone.