Turn fraud scores into decisions a reviewer can explain, within the number of cases a team can actually review.

A supervised XGBoost detector and an Isolation Forest verifier score each transaction. A capacity-aware policy blocks, allows, or routes disagreements to a person, with SHAP explanations on every decision.

On an IEEE-CIS temporal holdout, the capacity-capped detector showed 40.6% lower modeled loss than an amount-threshold baseline. That is an offline evaluation with modeled costs, not production savings.