Wait, Predict, or Defer? Evaluating an Offline Sequential Routing Policy for EEG Localization

Clinical systems need policies that determine when to act, collect more evidence, or defer to review. We study this decision through a retrospective sequential replay of 4,072 out-of-fold electroencephalography (EEG) segments from 157 patients. The agent receives one 10-second segment at a time and either issues a hemisphere localization, waits for another segment, or defers when it reaches its observation budget. Its primary routing signal, \emph{self-agreement}, measures the fraction of observed segments supporting the modal hemisphere prediction. We compare self-agreement with two other summaries of hemisphere votes, three model-output-magnitude signals, fixed evidence accumulation, and random routing. We select stopping thresholds using training patients and evaluate them on held-out patients across 20 repeated five-fold splits. At a target coverage of 40\%, self-agreement localizes (39.6\%\pm1.1\%) of patients with (0.886\pm0.013) accuracy after (7.51\pm0.66) segments. Vote margin and vote concentration achieve accuracies of (0.879\pm0.015) and (0.874\pm0.016), respectively, while the strongest output-magnitude policy achieves (0.714\pm0.025). Across target coverages of 20\% to 80\%, repeated-vote policies reduce selective risk by 0.086--0.248 relative to maximum output score. The advantage persists across observation budgets and randomized segment orders. These results establish repeated predictions as a stronger routing signal than model output magnitude in this replay. Prospective evaluation should measure how clinician review of deferred cases affects overall system performance.