Design the Monitoring Stack for a Churn Model

You are deploying a subscription-churn model that scores every active user nightly and hands the top 5% to a retention team. The label ("cancelled within 30 days of scoring") only becomes known 30 days after each prediction.

1. Lay out the monitoring layers you would build, from cheapest to most meaningful, with 2–3 concrete metrics per layer.
2. What would you use as a proxy for model quality during the 30-day window before labels mature?
3. How would you compute the "true" performance metric once labels arrive so that it lines up correctly with the drift signals?