Operational RF classifiers encounter signals outsidethe closed-set label taxonomy. We treat Unknown as a first-classoutcome via score thresholding on logits-derived confidence (maxprobability), predictive entropy, and energy score. We evaluatewith OSCR curves (Correct Classification Rate vs. UnknownFalse-Positive Rate) and AU-PR for unknown detection. Withlightweight hooks in the ensemble probability path, we realizerobust abstention with <0.1 ms … Continue reading Open-Set Handling in RF Ensembles: Thresholding, Abstention, and OSCR Analysis