Power from eight SEEG pairs before known response times gave 44.7 percent mean key agreement on 33 cue conflict trials in six recordings

We tested whether power from eight SEEG pairs follows the recorded key when it differs from an arrow cue. The reader used three bands and a half-second window before a supplied response event. Each of four source-order segments was held out. Fitting used only cue/key agreements in other segments. In OpenNeuro ds004473 v1.0.2, 33 valid conflicts occurred in six recordings. The reader matched 14 keys. Mean recording accuracy was 44.7 percent, with a descriptive 95 percent recording bootstrap interval of 20.8 to 71.1 percent. Mean probability assigned to the key was 0.498. This test did not establish key support rather than cue support. The best ordinary agreement decoder had no conflict trials. The package adds a conflict benchmark, a label identity proof, source code, all features and predictions, figures, exact splits, and raw retrieval records. Fresh checks of 12 pinned raw ranges rebuilt 18,540 power values with maximum log10 power error 2.15e-7. This is an offline test with known response times. It does not identify intention or test stimulation.

stereo-EEG · brain-computer interfaces · cue conflict · key decoding · power features · label leakage · OpenNeuro · cross-validation