Pulse variance correction gave AUC gains below 0.007 with intervals that included zero for the first five pulses in two SEEG recordings

We tested whether pulse variance correction improves response label ranking with five pulses. We used OpenNeuro ds004774 version 1.0.0. The fixed subset has 630 pulses at 63 surface stimulation sites, with six SEEG readout contacts in each of two recordings. Manual codes 1 and 2 were response labels. The selected data had 566 label entries. Each recording and rater was tested separately. The primary window was 10 to 90 ms. Both energy scores used the same baseline power scale. Correction increased first-five-pulse AUC by 0.002773, 0.002909, and 0.006039. All paired 95% site intervals included zero. Intervals that also resampled readout contacts included zero. Correction reduced the before-pulse energy floor. It did not establish a reliable response-label ranking gain. This result does not prove equal performance. The score uses a known cross-pulse product identity. Its unscaled unbiased expectation requires a common mean and independent zero-mean pulse errors. The normalized ratio has no such proof here. The package preserves native epochs, source metadata and labels, code, bootstrap arrays, four figures, licenses, and reproduction commands. A fresh check verified all 630 pinned raw ranges and all 5,034,960 SEEG voltage values. Independent checks passed. A new-folder run reproduced all saved tables, PNG figures, and bootstrap arrays. The evidence is conditional on these two recordings and source labels. It does not establish a neural source or BCI performance.

evoked responses · pulse variance · SEEG · brain-computer interface · AUC · energy score