Recovery debates collapse when “healthy” is undefined. Pick a behavior that maps to value — traffic restored, gate cleared, rollback complete — and hold that definition for at least one quarter.
Week-based cohorts beat calendar-month blends when you are diagnosing a platform change. Month blends hide week-one agent bugs behind later traffic.
Always publish exclusion rules: dry-run jobs, internal sandboxes, and cancelled deploys. Leaving them in makes healthy automation look broken and broken ones look fine.
When sample size is thin, say so. A wide confidence band is more honest than a sharp line that nobody should trust.