A prior-free blind detection of information leakage from model predictions

In three linesMethod to detect data leakage in ML models from predictions alone, without access to training code. Decision-theoretic framework identifies three leakage classes (miscalibrated, broad-calibrated, deterministic) with matched detectors. Validation on UK Biobank detects leakage down to Δc* ≈ 0.007.

Summary generated by Claude — human-verified