ICM2: Multivariate Conformal Convergence Diagnostics for Trustworthy Ensemble Predictions - A Conformal Prediction Perspective

Proposes ICM2, a multivariate conformal-prediction approach calibrating five convergence components (distributional agreement A, directional consensus D, uncertainty overlap U, perturbation invariance C, dependency penalty Pi) in R^5 via a quantile-envelope construction. The envelope is coupled with a two-stage MondrianICM conformal calibration, and each ACT/DEFER/AUDIT decision is decomposed through Shapley values over 32 component coalitions. ICM2 calibrates in the diagnostic space rather than the label space, yielding per-component interpretability without requiring an assumed covariance structure or joint-score model. Across 20 classification datasets, MondrianICM is on par with SAPS on prediction-set size (mean 1.200 vs 1.194 at alpha=0.10) while adding Shapley attribution. CIFAR-100 collapse with homogeneous 5xResNet-20 (19.9) is reported honestly; diverse three-architecture ensemble achieves 2.548 +/- 0.185.