Uncertainty-Driven Selective Prediction with Auditable Human Deferral for EEG Decoding of Participant-Rated Occupational Risk

Large-language-model (LLM) layers are increasingly appended to electroencephalography (EEG) decoders, although their decision contribution is often unclear. We evaluate an uncertainty-driven architecture in which a calibrated EEG classifier predicts participants’ own four-level occupational-risk keypress ratings, a deterministic rule either retains or defers a prediction, and a template or optional LLM renders an auditable report. The objective is not to replace a previously published classifier benchmark on the shared EEG archive, but to evaluate downstream decisions from a fixed frozen-representation decoder. The analysis used 2634 response-inclusive 3 s epochs from 31 retained workers. Subject-grouped three-fold evaluation yielded balanced accuracy of 0.6336 (95% CI 0.5745–0.6622) and accuracy of 0.6875. Peak probability distinguished correct from incorrect predictions only moderately (AUC 0.6529), and the coverage-risk relationship was non-monotonic. At tau = 0.5, 71.49% of trials were retained (balanced accuracy of 0.6829); with illustrative under-estimation/over-estimation/deferral costs of 5/0.5/1, normalized cost fell from 0.5816 to 0.5363, but sensitivity analysis showed that benefit depends on the cost setting and threshold. A rule engine reproduced all evaluated LLM risk/deferral decisions, while a deterministic template matched the prespecified reporting checks; therefore, no decision-performance advantage of the LLM was demonstrated. The study supports cross-participant prediction within the shared experimental setting, not stimulus-disjoint generalization or objective workplace-risk prediction. Because response timestamps and actual supervisor decisions were unavailable, motor/visual confounding cannot be excluded and the human handoff remains a simulated policy rather than a human–AI team evaluation.