A 10% conditional error test stopped five of eight fixed SEEG readers in an offline replay with complete trial feedback

We tested a feedback rule for a fixed brain state reader. We reused eight public SEEG recordings from OpenNeuro ds004473 version 1.0.2. Each trial had two selected windows. We froze the readers and their confidence gates. We then tested a 10% conditional accepted-output error target in the next task condition. One update used the accepted count and wrong count from the whole trial. With complete feedback before the next trial, the rule stopped five of eight readers. The offline stop replay withheld 149 of 248 wrong accepted outputs and 249 correct outputs. It still delivered 99 errors. Two readers had observed error above 10% and did not alarm. The ideal rule needs an explicit per-trial conditional error assumption. Under that assumption, the chance of any false alarm across eight readers is at most 5%, at any time. In a separate null simulation with duplicate errors, separate window updates gave 19.96% alarms under a nominal 5% rule. Trial updates gave 3.01%. This is an exploratory reuse of selected event windows. No alarm is not a risk certificate. The package has a proof, reusable code, figures, all feature inputs, fixed models, source receipts, and independent checks. It does not test live command control or brain stimulation.

sequential monitoring · conditional error rate · brain-computer interface · stereoelectroencephalography · selective classification · false alarm control