Full-range Binary Classifier Calibration for Stable Model Updates in Production

This paper introduces a calibration method for binary detection models whose malicious distribution changes rapidly while the benign distribution remains comparatively stable. Instead of calibrating class probabilities, it targets the entire false-positive-rate (FPR) curve so that scores retain a consistent operational meaning after retraining and redeployment. On one held-out split, reported relative FPR error was at most 2.3% from 10% down to 0.1% FPR, and 7.2% at 0.01% FPR. The shipped calibration artifact stayed below 200 KB across calibration sets ranging from 1,000 to 10 million benign samples.