Pair choice with five reused controls raised background calls under random roles in eight SEEG sessions

We tested energy alarms with public SEEG data from OpenNeuro ds004696 version 1.0.1. The test used eight epilepsy sessions, 86 stimulation sites, and 886 candidate recording pairs. Five early after windows defined each pair response score. The alarm reused five background windows for pair choice and energy rank. We tested all random target roles within six-window subsets of ten recorded before windows. The response window was 20 to 500 ms after a pulse. The background window was 980 to 500 ms before a pulse. The random-role test gave 20.82 percent background calls with reused controls. A symmetric pair choice gave 16.67 percent. It includes the new vector in pair choice and needs every candidate pair. We prove a conservative rank bound when full background vectors are exchangeable conditional on the response scores. The proof does not establish exchangeability of chronological SEEG windows. On the actual later windows, background calls fell from 20.61 to 16.25 percent. After-window calls also fell from 84.11 to 78.88 percent. The package preserves the proof, monitor code, data products, source hashes, all decisions, figures, and reproduction instructions. This is an offline calibration study. It does not establish a neural response, a live BCI benefit, or a 5 percent alarm rate.

rank calibration · selection bias · conformal prediction · stereoelectroencephalography · SEEG · brain-computer interface · electrical stimulation · energy monitor