- Load two or more normalized snapshots with compatible account, timezone, currency, metric, and attribution definitions.
- Validate data freshness and finalization windows before comparing periods.
- Separate expected learning, seasonality, reporting latency, and planned changes from unexplained anomalies.
- Evaluate pacing, delivery, conversion quality, unit economics, creative fatigue, tracking health, policy status, and changed account objects.
- Return observations, confidence, likely causes, required investigation, and decision thresholds. Do not mutate the account.
- Persist a versioned monitoring bundle and link detected failures to regression or follow-up tasks.
Do not alert on percentage changes with trivial denominators or incomparable windows. State when evidence cannot distinguish noise from a material change.