Detection models in adversarial environments encounter rapidly drifting malicious distributions while benign distributions remain stable, necessitating teams to retrain and redeploy frequently to counter new threats. Retraining often alters output prediction scores, disrupting downstream users of the model. For these security-focused models, a consistent false-positive rate (FPR) across all output values is essential, whereas standard probability calibration methods focus on class probabilities rather than an FPR contract.
We introduce a method built on existing calibration primitives that targets the entire FPR curve, ensuring consistent FPR meanings across deployments. On a held-out split, the observed relative FPR error was at most 2.3% from 10% down to 0.1% FPR and 7.2% at 0.01% FPR. The shipped artifact remains under 200 KB in measurements across calibration sets from 1K to 10M benign samples.