Full confidence fitting lowered mean Brier score by 0.00016 against an offset update for joint 4–10 and 4–8 Hz phase success in seven later SEEG rest records

We tested confidence in fixed SEEG phase forecasts from 28 contact pairs in seven people. The data came from OpenNeuro ds006065 1.0.0. Main success required error at most 45 degrees against both offline 4–10 Hz and 4–8 Hz references. Inputs used only past voltage. The target was about 50 ms after the input cutoff. Both confidence updates used earlier rest. The test used separate later rest. Full fitting lowered mean Brier score from 0.223832 to 0.223670 against an offset update. The mean gain was 0.000162. The person bootstrap 95% interval was -0.000029 to 0.000358. The sign-flip p value was 0.203. This does not establish an added mean benefit. The offset update reduced mean joint-event probability error from +8.79 to -0.10 percentage points. It preserved every quantile gate decision. A secondary 5–9 Hz test with the same band center gave a larger slope-fitting gain. The package preserves code, voltage inputs, fixed and new models, joint labels, two figures, results, and retrieval instructions. Full public source hashes and every retained sample were checked. The cohort is reused. The labels need future voltage. We did not test live stimulation, biological phase truth, or clinical benefit.

stereoelectroencephalography · theta phase · phase forecasting · confidence calibration · Brier score · brain-computer interfaces · offline phase reference