A 25% random trial audit rate and a delay of five trials increased mean error exposure by 23.9 points against complete immediate feedback in eight offline SEEG tests

We tested how random trial audits and delayed labels affect a fixed SEEG error monitor. We used public OpenNeuro ds004473 version 1.0.2, eight reused recordings, 975 next-condition trials, and 961 accepted outputs with 248 errors. The readers and gates stayed fixed. The main setting used a 25% independent whole-trial audit rate and a delay of five retained trials. Complete immediate feedback was the reference. Mean person error exposure increased from 63.3% to 87.2%, a change of 23.9 percentage points. The descriptive 95% whole-person bootstrap interval was 9.5 to 37.9 points. Mask Monte Carlo standard error was 0.052 points. Across the fixed cohort, expected wrong delivered outputs increased from 99 to 184.4. Trial label requests fell from 574 to 208.2. Correct outputs withheld fell from 249 to 82.8. We scored 10,000 saved masks per person at five audit rates and three delays. A proof shows that independent audits and ordered fixed-delay feedback preserve the conditional false alarm bound for one fixed setting per reader. The bound needs a conditional count null and correct labels. No alarm does not certify low risk. All 162 source input files matched the sci manifest. Eight fresh pinned EDF ranges matched their hashes. Independent scalar checks passed in 4,200 cases, and a full numerical replay matched 21 files. The package preserves source features, raw retrieval recipes, code, masks, results, proof, three figures, and reproduction commands. This offline test does not establish a population effect, live device performance, or a brain stimulation result.

stereoelectroencephalography · brain-computer interface · error monitoring · random audits · delayed feedback · selective classification · conditional error bound