CSRE-MOL: CONDITIONAL RESIDUAL EVIDENCE FOR MOLECULAR PROPERTY PREDICTION

A frozen molecular classifier supplies a global score, while labeled analogues expose an inference-time structural signal that may repeat what the classifier already knows. We ask a sharper fusion question: which part of retrieval evidence remains after accounting for the score being augmented and for neighborhood support? We present Calibrated Structural Residual Evidence for Molecular Prediction (CSRE-Mol), a post-processor that estimates the predictable component of normalized retrieval evidence with a task-wise ridge projection and applies only the conditional residual to the frozen score. A positive-slope Platt map then separates score correction from probability calibration. On the existing development benchmark of GIN, GraphGPS, GraphMVP, and Uni-Mol across eight MoleculeNet classification datasets using three random seeds, CSRE-Mol raises mean ROC-AUC from 70.07% to 75.97%, mean average precision from 45.81% to 52.99%, and lowers mean NLL from 0.3933 to 0.3625. The predictable evidence component alone reaches 69.85% ROC-AUC, whereas the conditional residual reaches 75.97%; a matched conditional permutation falls to 70.48%. These controls support a conditional-innovation interpretation of the gain rather than treating retrieval as a second predictor. The evaluation uses an existing development split, and Uni-Mol/ToxCast identifies a useful boundary condition: conditional correction helps when the residual neighborhood signal is informative, but cannot recover ranking signal that is absent.