Shadow deployment for fraud models

Design a shadow-deployment and validation harness for promoting a new fraud-scoring model to production, where you can't just A/B it freely because a worse model means real fraud losses or wrongly-blocked legitimate payments, both costly. You need confidence the new model behaves correctly on live traffic before it makes any real decision. Walk through how shadow mode works, what you compare to gain confidence, and how you close the gap between offline-eval results and live behavior.