Testing the Test: Score-Direction Instability in Class-Split Anomaly Detection

In three linesAuthors show that class-split evaluation in anomaly detection becomes ill-posed when the held-out anomaly class overlaps the normal mixture in representation space. Anomaly scores may collapse or invert. A simple training-free diagnostic (neighborhood class leakage) predicts score-direction instability across Fashion-MNIST, CIFAR-10, Imagenette.

Summary generated by Claude — human-verified