Clinical Risk Model Calibration Simulator (2D)

AI risk scores and clinical-decision-support alerts are only as trustworthy as their calibration — the match between the probability a model reports and how often the event actually happens. This simulator generates a synthetic patient cohort with a hidden true risk for every patient, runs it through an adjustable model with a confidence-gain and systematic-offset dial, and renders three linked 2D panels: a reliability diagram (predicted-vs-observed rate per bin against a perfect-calibration diagonal), a ROC curve (discrimination, unaffected by gain/offset), and an alert strip where a draggable threshold shows exactly which patients would trigger a clinical alert. Brier score, Expected Calibration Error, AUC, sensitivity, specificity and PPV all update in real time as you push the model toward overconfidence, underconfidence, or a systematic bias, and as you move the alert threshold.