In 65 error trials, the probe IN/OUT score resolved neither response nor memory content in 14 SEEG subjects

We tested if a probe IN/OUT decoder reads the response or the true probe class. Error trials separate the two. We used the archived band-power features of 314447a0 (ds004752, ses-01). We trained shrinkage LDA on correct trials only and scored 65 error trials in 14 subjects. The plan was written before the analysis. In the primary window (0.5-1.0 s after the probe), the error-trial score was near the decision midpoint: mean signed score -0.11 SD toward the response (95% CI -0.43 to 0.18; 7 of 14 subjects positive; exact sign-flip p = 0.53; pooled AUC for the response 0.50, permutation p = 0.97). Correct trials gave +0.32 SD. Post hoc, the error-trial score was lower than the correct-trial score (difference 0.38 SD, 11 of 14 subjects, p = 0.035, not corrected). Thus the decoder did not follow the reported choice on error trials as it does on correct trials. A simple motor-plan signal does not explain this. The data cannot reject a truth-following score of the same size. Limits: few error trials, one session, response inferred from Correct and Match, button mapping not checked.

intracranial EEG · SEEG · verbal working memory · error trials · shrinkage LDA · brain-computer interface · decision evidence