Student t scores reduced mean probability loss with simulated fourfold voltage gain in five SEEG tests at 100–500 ms

We tested probabilities for known stimulation sites in public OpenNeuro ds004457 version 1.0.2. The test used 138 sites, 15 readouts, and five people. Three pulses fitted each template. A separate fourth pulse fitted the probability scale. The last four pulses supplied the test mean. The main window was 100 to 500 ms. Only test voltages received a fixed simulated fourfold gain. Student t scores reduced equal-person mean log loss from 5.4392 to 4.0838 bits per site test, compared with Gaussian scores. Both used finite template scaling. Loss was lower in four of five people. The descriptive person bootstrap range for the mean difference was -2.0250 to -0.5461 bits. The sign-flip reference fraction was 0.125. Student t made no decisions at 90% probability under fourfold gain. It removed correct confident decisions as well as wrong ones. At original voltage scale, its mean loss was 0.0502 bits higher. All readers predicted the same site at each setting. A proof gives a distant-input confidence bound under a common metric, equal priors, and fixed positive scale. All 110 raw source files passed hash checks. A raw rebuild matched 2,453,000 voltage bins and all event fields. The package includes code, compact data, fitted readers, all probabilities, three figures, source hashes, and reproduction steps. This offline test does not establish live BCI performance or writing safety.

stereoelectroencephalography · site classification · Student t scores · probability loss · voltage gain · confidence coverage · brain-computer interfaces