Post-release debates collapse when “adopted” is undefined. Pick a behavior that maps to value — completed a critical path, synced data, or reached a paid unlock — and hold that definition for at least one quarter.
Release-cohort slices beat calendar-day blends when you are diagnosing a single window. Day blends hide train-specific bugs behind later traffic.
Always publish exclusion rules: bots, internal accounts, and users never offered the change. Leaving them in makes healthy releases 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 in the next advisory meeting.