Three consecutive SEEG scores detected 14 of 320 later key responses after early threshold calibration in eight patients

We tested whether three consecutive decoder scores change command detection after early threshold setting. We used public SEEG data from eight patients in OpenNeuro ds004473 version 1.0.2. We kept the archived model, score grid, and chronological split. Each rule had its own threshold set on 20 earlier responses per patient. The early budget was at most one false alarm per non-target minute. In 320 later responses, three scores detected 14 responses before the key press. One score detected 31. Mean later false alarm rates were 1.89 and 1.66 per non-target minute. The planned patient sign-flip test gave p = 0.125. It did not establish a mean change under its assumptions. The result concerns this fixed offline benchmark. It does not establish equal performance or an online control effect. The package contains a causal alarm method, patient results, figures, source receipts, code, and instructions. A fresh check rebuilt all 34,268 continuous scores from 769 version-fixed public source ranges. Saved training features were used; whole EDF hashes and all raw training windows were not checked again.

brain-computer interface · stereoelectroencephalography · neural decoding · alarm calibration · temporal persistence · false alarm rate · offline benchmark