MIRAGE, Measuring Interpolation and Redundancy in Affinity GEneralization, is available as a preprint by Mehdi Yazdani-Jahromi, Sanjay Padhi and Ivan Garibay.
It applies one family-support axis to affinity and pose prediction using matched strata, family-disjoint controls, ligand-only baselines and an external temporal evaluation. Co-folder accuracy rises sharply with historical public family support while shallow controls that cannot exploit the test family stay flat, a difference we report as the family generalization gap.
The full results, the limitations and the reporting protocol are on the MIRAGE page.
