A confidence fit across ±9.77 ms gave no clear group gain in mean maximum SEEG log loss in eight patients

This offline test asked whether confidence fitting across time offsets reduces maximum log loss for a SEEG stimulation-condition reader. It used public STReEF data, eight patients, four conditions per patient, and 160 held-out pulses. Three pulses per condition fitted the reader. Two separate pulses fitted one inverse temperature. Each test pulse had 41 offsets across ±9.77 ms. Fitting each calibration pulse maximum loss reduced mean maximum test loss from 15.761 to 1.056 nats relative to fitting at recorded times. The patient bootstrap interval for the difference was [-43.236, +0.008]; the sign-flip value was 0.1875. One patient accounted for 97% of the mean decrease. The decrease depended on the selected temperature limit. Only 40 test pulses occurred after fitting all conditions. That subset had no loss benefit and contained one condition per patient. The evidence gives no clear group or real-time BCI gain. The package adds a finite-set convex fit, a calibration-set bound, a failure audit, native voltage extracts, source hashes, code, three figures, and repeat instructions. Fresh public source recovery and the full numeric repeat passed.

SEEG · confidence calibration · time error · log loss · brain-computer interface · stimulation · temperature scaling