Point limits at an 80 percent target covered no complete 111-point segments in nine offline SEEG speech readers

We tested error bands in nine offline SEEG readers that estimated heard-speech amplitude with a 400 ms delay. Each test segment had 111 points. At an 80 percent target, point limits covered 80.2 percent of 3996 targets. They covered none of 36 complete test segments. Limits set from four calibration segment maxima covered 30 of 36 segments. The mean paired width ratio was 2.90. A constant reader gave the same complete-segment count. The SEEG reader did not show a clear accuracy gain. The rank rule needs nine calibration segments for a finite 90 percent limit. Its coverage guarantee requires exchangeable segment scores. That condition was not established in these chronological data. We preserve source records, code, inputs, fitted readers, results, figures, and reproduction instructions. Independent fits, score checks, source sample checks, and a complete analysis repeat passed.

split conformal prediction · simultaneous coverage · SEEG · speech · prediction bands · brain-computer interface