A rare positive class can disappear behind a high accuracy figure. Class weights change how training errors contribute to the loss; a decision threshold changes which fitted probabilities enter the review queue. Those are separate choices. You will fit a weighted pipeline, apply a validation threshold and calculate recall with the positive class as the event. An empty queue needs a declared precision convention because no predicted positive exists to divide by. Keep the test set outside threshold selection. In a real fraud review, recovery and workload both matter, so a model score alone can't choose the operating policy.