A fixed SEEG confidence score gave no clear gain in later offline error forecasts in 21 MetaRDK recordings

We tested whether a fixed SEEG score trained on reported confidence improves later choice error forecasts. The test used public MetaRDK data from 21 recordings. We trained the reader on early trials, fit error probabilities on a separate middle block, and scored 2,274 later trials. Adding the neural score to six task inputs changed equal-person mean Brier score by -0.00116. The descriptive 95 percent person interval was -0.00430 to 0.00156. Eleven people improved. This test did not establish a clear gain or a zero effect. A constant error-rate forecast had less error than the task model. In an exploratory check, removing trial number reduced task-model error. Adding the neural score to that reference changed Brier score by +0.00030. The package supplies a time-ordered confidence-to-error test, compact data, all fitted models and forecasts, three figures, checked source records, and code. One full source signal was rebuilt exactly. Source processing prevents a real-time claim. No interface or stimulation outcome was tested. The cohort overlaps earlier studies.

SEEG · confidence · choice error · error forecasting · Brier score · brain-computer interface · MetaRDK