Limits set for 80% sample coverage covered zero of 36 complete test segments in nine offline SEEG readers

We tested whether limits for individual sound targets covered complete speech segments. We used public SEEG data from OpenNeuro ds004703 version 1.1.0. Nine people each had two fit segments, four calibration segments, and four later test segments. Each segment had 111 consecutive 100 ms sound targets. An eight-pair SEEG reader used a 400 ms delay. Limits set for 80% sample coverage covered 80.2% of targets but zero of 36 complete test segments. Calibration with segment maximum errors covered 30 of 36 segments. The mean within-person width ratio was 2.90. A constant sound-level mean also covered 30 segments. The reader had no mean width or prediction-error advantage over that baseline. The original archive includes code, inputs, figures, source records, and a rank-limit function. SCI retains the code, figures, source records, and results; incompatible dataset inputs are omitted. The rank proof requires exchangeable segment scores. Ordered recordings do not establish that assumption. These results apply to short passive-listening segments and an offline reader.

stereoelectroencephalography · conformal prediction · simultaneous coverage · brain-computer interface · speech decoding · uncertainty calibration