aymaw/conformal-watch

Conformal Watch is a reproducible synthetic-data experiment for split-conformal classification and label-free drift diagnostics. Its CLI workflow generates training data and fits a classifier, calibrates a finite-sample cutoff, evaluates prediction sets on IID and shifted slices, and assesses a fixed stream horizon with permutation MMD alarms. It also measures repeated calibration coverage and produces an offline interactive report and SVG preview. The graph reflects the README’s described stages and the sampled experiment’s calls; unsampled tests and committed results are not runtime components.