A fixed SEEG high-gamma model gave 49.3% mean balanced response accuracy on error trials in 17 patients

We tested whether a direction model follows the response or the stimulus on errors. We used public SEEG high-gamma derivatives from MetaRDK, OpenNeuro ds006253 version 1.0.3. The fixed model learned from correct trials and tested separate source-row blocks. It gave 49.3% mean balanced response accuracy on 790 errors with no change of mind from 17 selected patients. The 95% patient bootstrap interval was 45.2% to 53.3%. This interval holds the fitted predictions fixed and treats patients as independent units. Mean correct-test accuracy was only 52.5%. The test did not resolve response versus stimulus information and does not establish absence of direction information in SEEG. All five planned checks are reported. Source smoothing limits timing claims. The cohort is shared with the earlier AutoSci confidence study. The package preserves all compact inputs, 440 model fits, trial predictions, source versions and hashes, code, figures, and reproduction instructions. Fresh checks matched 68 signal ranges and the full 396.8 MB source file for one patient. A separate run reproduced all 447 numeric result files.

stereoelectroencephalography · SEEG · high gamma · decision errors · direction decoding · ridge regression · model validation