Design a fraud-detection system with heavy class imbalance and an adversarial opponent.

Combine fast deterministic rules with learned risk scores and a review path. Optimize at the operating capacity, use time-based backtests and adversarial slices, log decisions, and support rapid rule or model rollback.

Make interfaces and ownership explicit; add versioning, access control, monitoring, canary rollout, rollback, and a plan for delayed labels or human review.