Extend post-fit repair objectives and preservation scope

Develop richer post-fit objectives, alternative selection principles for choosing among exactly prediction-preserving repaired probability reports, and extensions of CORD beyond top-1 prediction preservation.

Background

CORD establishes a post-fit approach for preserving the original classifier’s top-1 prediction while modifying the calibrated probability vector. The repair is selected using a particular Bernoulli–KL objective, a local-reference rule, and an inherited feasible-mean constraint. The authors explicitly identify broader repair objectives, different principles for selecting among feasible repaired reports, and preservation guarantees beyond the top-1 decision as unresolved directions for future research.