Class-conditional robust feature geometry
=========================================
Classes: 3
Training features per class: 220
Feature dimension: 12
Class-wise leverage contamination: 10%

Method                              AUROC    mean in-score    mean OOD-score
Empirical class covariance         0.481            3.317            3.267
Robust FastMCD class geometry      1.000            3.690            9.599

Interpretation
--------------
The score is the distance from each test feature to its nearest fitted class geometry.
Empirical class covariance is weakened when each class reference set contains leverage-like contamination.
Robust class-conditional geometry estimates each central class shape more stably, which improves nearest-class OOD separation in this synthetic example.
