Early target limits left negative mean error skill in ten SEEG readers of heard sound amplitude at zero added delay

We tested fixed early target limits on ten stereoelectroencephalography (SEEG) readers of heard log sound amplitude using 95 selected passages from OpenNeuro ds004703 version 1.1.0. At zero added delay, the limits raised equal-person mean error skill from -57.2% to -53.2%, a gain of 3.97 percentage points. The descriptive 95% whole-person bootstrap interval was 0.00 to 11.78 points; the exact sign-flip p value was 0.50. Two people improved and eight had no change. One person supplied 98.3% of the gain. The limits changed 36 of 4446 later outputs, with 99.66% mean target coverage; no primary row had more squared error, but large errors remained inside the range. Narrower early limits increased error at three target times in the 400 ms secondary test. A proof gives the condition under which projection protects squared error. This offline analysis reuses the same cohort and later passages. It does not test brain stimulation or live BCI control. Independent fits, score checks, a second offline run, and a fresh-source rebuild passed. Source recordings, event metadata, and derived input caches are excluded because the dataset metadata says CC0 while its README states noncommercial and no-identification conditions. The package retains the results, figures, code, source URLs and hashes; reproducing the analysis requires rebuilding omitted inputs under source terms.

stereoelectroencephalography · brain-computer interface · ridge regression · output limits · error skill · squared error · passive listening