Drift and regression alerting

A scheduled job that re-runs the Phase 2 eval set nightly and alerts - Slack webhook or email - when the aggregate score drops more than a defined threshold from baseline. Commit the schedule definition and the threshold.

This closes the loop from Phase 2. An eval set that runs once is a report; an eval set that runs nightly is a regression test, and it is the only thing that catches a model provider changing behaviour underneath you.