Task history reduced prediction loss for three memory loads in test SEEG sessions from 14 patients at penalty 10

We tested task history and hippocampal power as predictors of four, six, and eight letters in public SEEG data. We used OpenNeuro ds004752 version 1.0.1 and exact reused feature arrays from 14 patients with epilepsy. Separate multinomial logistic models trained on the earlier half of numeric session labels and tested on the remaining sessions. The main penalty was 10. There were 1,312 training trials and 1,506 test trials. History lowered mean patient test log loss by 0.048898. The 95 percent patient bootstrap interval was 0.021955 to 0.071156, and Holm-adjusted sign-flip p was 0.009277. Thirteen patients had lower loss. Adding six mean hippocampal power bands during late maintenance increased loss by 0.088680 at penalty 10. At penalty 100, the interval included zero. The source task audit found five exceptions to the stated after-error load rule; no labels were repaired. This exploratory study supplies a control model and a reproducible probability benchmark. It does not show that hippocampal load information is absent. The package preserves source pins, exact inputs, code, fitted parameters, predictions, patient scores, bootstrap draws, raw checks, and repeat instructions. All 62 sessions passed selected fresh raw-window checks. Three full sessions passed exact feature rebuilds. All 34 numerical outputs matched an offline repeat.

SEEG · memory load · task history · hippocampus · log loss · multinomial logistic regression · cross-session prediction