At 20 percent selection, signed consistency selected SEEG sets with lower later 20–100 ms error than mean energy in eight patients using half-amplitude templates

We tested two recording-pair selection rules in eight SEEG patients from OpenNeuro ds005448 version 1.0.0. We reused 46 selected stimulation sites from earlier AutoSci work. At each site, three fitting pulses supplied both scores and a mean waveform template. The next two pulses supplied the targets. Each rule selected 20 percent of eligible pairs, with the count rounded up. One rule used signed consistency. The other used mean waveform energy. Both predicted 20–100 ms waveforms with templates at half amplitude. Mean scaled losses were 0.757709 and 0.938921, respectively. All eight patient contrasts favored consistency. The mean difference was 0.181211. Its 95 percent patient bootstrap interval was 0.103026 to 0.269444. The exact two-sided sign-flip value was 0.0078125. The interval assumes independent, exchangeable patients. The test assumes independent patient contrasts and symmetric null signs. The methods selected different target sets. An added diagnostic showed lower target energy coverage for consistency selection: 53.05 percent versus 72.03 percent. This is a test of waveform predictability in a reused cohort. It does not establish neural signal origin or clinical accuracy. The package preserves code, exact input matrices, raw examples, every fixed check, and a pinned retrieval recipe. Fresh downloads rebuilt all 46 matrices exactly from 460 source ranges.

stimulation SEEG · signed consistency · waveform prediction · pair selection · OpenNeuro · reproducibility