A noise rule gave a mean error estimate below the mean test error under random pulse orders in eight SEEG recordings

We tested a noise rule that stops SEEG calibration after three to five pulses. The test used public OpenNeuro ds004696 version 1.0.1, with 317 inherited conditions from eight people. Under uniform random orders of the recorded pulses, mean test error exceeded its estimate by 0.00794 after division by fixed pulse energy. The descriptive 95% person bootstrap interval was [0.00422, 0.01262]. The difference was positive in all eight people. The rule used a mean of 4.671 pulses. For each fixed count, the expected difference is exactly zero; the proof does not require independent source pulses. We used all 30,240 first-five orders for 179 conditions and 20,000 draws for each other condition. The recorded-order result was different: the mean difference was -0.04848, with interval [-0.10922, 0.00200]. The random-order result does not establish a low estimate in live time order. Raw checks from 16 channels matched 1,164 matrices exactly. A separate repeat matched all numeric tables. The package preserves figures, inputs, code, proof, plan, source hashes, retrieval recipes, and instructions. This is a new error-estimate audit of a reused cohort. It does not test BCI task performance or stimulation control.

SEEG · brain-computer interface · stimulation calibration · error estimation · stopping rule · randomization