Multi-Level Adaptation of Rough Inclusion Classifiers with Feature and Pairwise Corrections

How much structure should be added to a rough inclusion classifier when its original similarity score is insufficient? We study a multi-level adaptation procedure that uses training validation to select among the unmodified classifier, score calibration, univariate value-dependent corrections, and constrained pairwise correction tables. The complete base-tuning procedure is repeated within cross-fitting to learn and evaluate these extensions. Across 240 train/test splits of eight binary datasets, macro-averaged balanced accuracy increases from 82.05% to 85.93%. Almost all of this improvement comes from calibration and univariate effects: allowing pairwise corrections adds only 0.02 percentage points over the selected univariate control. Although the pairwise mechanism reliably recovers strong synthetic XOR interactions, validation gains only weakly predict test gains on real datasets. An EBM correction layer initialized with the same base scores remains competitive, limiting claims of a distinctive performance advantage. These findings separate the ability to represent interactions from the ability to select interactions that generalize. Our contribution is an auditable adaptation of this classifier family and an empirical analysis of when additional correction levels help. Within the studied setting, the results identify reliable structure selection, rather than additional pairwise representational capacity, as the more pressing challenge.