A 100 ms wait did not keep causal 70 to 110 Hz power error below 5% of earlier power after holding voltage across 20 ms gaps in 35 SEEG records

We tested known sample loss in four fixed channels per record from public OpenNeuro ds006519 version 1.0.0. The main test used 20 ms synthetic gaps, exact prior filter state, held voltage through the gap, and causal 70 to 110 Hz power with a past 100 ms mean. At 100 ms after gap end, error exceeded 5% of earlier power in 346 of 408 cases from 35 records and 21 patients. The rate was 85.2% with equal patient weight. We give a proof and code for a conditional power interval under a specified missing-voltage bound. All observed fills met that bound, and the interval covered every evaluated sample in 4944 tested cases within the numerical tolerance. Its midpoint reduced the main paired bounded delay by 1.01 ms, with a descriptive patient bootstrap interval of 0.81 to 1.23 ms. The midpoint had much larger actual error before eligibility. Other tested conditions showed later increases above the error limit after a first eligible sample. The package preserves exact selected voltage, fresh source checks, code, figures, all scores, and reproduction instructions. This is a software gap benchmark. It does not establish neural truth, decoder gains, or clinical read-after-write safety.

SEEG · causal signal processing · sample loss · power feature · conditional error interval · brain-computer interface