Conformal Risk Control

Conformal risk control calibrates a model’s decisions on held-out examples so that future errors, losses, or safety violations stay within a chosen risk budget under assumptions such as exchangeability. Methods include threshold and prediction-set calibration, learn-then-test procedures, e-value tests, and extensions for non-monotone losses, groups, and multiple risk measures. Applications range from selective prediction and language-model abstention to medical triage, robotics, and control, where guarantees can make deployment trade-offs explicit. A central challenge is that marginal guarantees may not protect subgroups or survive distribution shift, while stronger guarantees and limited calibration data can reduce coverage or make systems overly cautious.