Separate confidence updates for four pulse phases increased mean Brier score in seven later SEEG rest records with a 1 s holdoff

We tested confidence for fixed 50 ms phase forecasts at simulated pulse times. The data were 28 contact pairs with SEEG labels in the source metadata from seven people in OpenNeuro ds006065. We kept the phase readers and confidence slopes fixed. We used 56 s of earlier rest to fit new confidence intercepts. At a 1 s holdoff, separate updates for four pulse phases increased later mean Brier score from 0.22757 to 0.22878, compared with one update from regular times. Five people had an increase. The descriptive 95 percent interval for the signed gain was -0.00357 to 0.00109. The mean signed bias decreased, but total probability error increased. Planned checks gave a mean gain at a 0.1 s holdoff and a mean loss at a 2 s holdoff. A post hoc test kept the later 0.1 s schedule fixed. Using all 235 to 386 early examples per phase gave a mean gain of 0.00164. Using 50 sampled examples gave a mean loss of 0.00183 across 20 subset draws. These results support checks of confidence at selected output times and of the available calibration data. They do not establish a general minimum count. The phase label uses a centered filter and future voltage. No pulse changed the voltage. The package includes selected voltage, fixed models, event indices, fit plans, code, source records, results, four figures, and exact replay checks.

SEEG · phase forecast · confidence calibration · Brier score · closed-loop brain interface · theta phase · trigger timing