Eight high-gamma means increased mean offline forecast loss for later changes of mind in six MetaRDK SEEG sessions under a fixed ridge model

We tested forecasts of later recorded changes of mind in six public MetaRDK SEEG sessions. The test used 500 later trials with 47 changes. It conditioned on no recorded change through 150 ms after movement onset. The task model used stimulus, decision time, and preceding task history. Adding eight high-gamma means raised equal-person mean log loss from 0.3333 to 0.3871 natural-log units per trial. The fixed ridge penalty was 10. Four of six persons had higher loss. The primary gain was -0.0538, with a descriptive person-bootstrap interval of [-0.0867,-0.0200]. The exact two-sided person sign-flip value was 0.125. Fixed penalties 1 and 100 also gave negative mean gains. The constant fitting-frequency model had lower mean loss than the task model. An audit found 24 of 193 source changes at or before recorded movement onset. The public smoother and session normalization prevent a live-performance claim. The result does not show that the brain lacks a useful change signal. The package preserves code, trial records, model states, three figures, metadata, and source retrieval instructions. All six public signal files were retrieved again. Rebuilt features matched byte for byte. Independent metrics and 72 model refits passed. A second run reproduced every numerical output exactly.

stereoelectroencephalography · high-gamma activity · change of mind · forecasting · logistic regression · offline evaluation