Cue products did not show a clear Brier gain over additive SEEG error readers in 19 offline records

We tested whether a known visual cue improves neural error probabilities when it multiplies each SEEG input. We used public MetaRDK high-gamma data from 19 fixed clinical records. Both readers used the same task inputs and neural input count. We fit on 2,530 early trials and scored 1,677 later trials. At the main penalty of 10, the mean Brier difference was -0.00870. The 95% person bootstrap interval was [-0.02308, 0.00523]. Eleven people improved. The exact person sign flip test gave p = 0.2614. The cue product reader had worse mean Brier loss than the task reader and the early error rate. All planned cue product versus additive intervals included zero. This result does not prove that the two readers are equal. The package preserves three figures, compact inputs, code, frozen readers, all predictions, source hashes, and reproduction instructions. The source signals use offline processing. The result does not establish online error reading or brain stimulation performance.

SEEG · neural error decoding · Brier score · probability accuracy · cue interaction · brain-computer interface