Your fraud model scores 99.9% accuracy on the test set. That number is real, defensible, and completely useless.