Two response comparisons gave mean four-group accuracy above chance for fixed sounds in 20 iEEG patients and 16 fMRI adults

We tested a small reader for four broad sound source groups in separate public human studies. The input was a signed iEEG LFP mean or a BOLD fMRI estimate. Each prediction used two relative response comparisons from three trained pairs. At most six sites were needed after calibration. Four people per study supplied development. A local code lock preceded reserved tests in 20 patients and 16 MRI adults. Mean balanced accuracies were 32.80% and 48.98% against 25% chance. Their 95% person Student intervals were [28.08, 37.51]% and [45.26, 52.70]%. The shared mean rule passed, p=0.00182343. Common label references repeated pair fitting and development tree choice. A later arithmetic repair made the batch reference agree with the locked scalar reader; all initial outputs remain. Six electrical people were below chance. Electrical Animal and Other recalls were below 25%, so the result does not confirm success for every group. The small reader lost 15.26 fMRI accuracy points against a reader that used all candidate site ranks. Joint person and sound resampling, with pair refits, gave conditional 95% ranges of 26.37–39.68% and 38.29–53.33%. An added check with six candidate sites retained 17 patients and 16 MRI adults. Both mean accuracy intervals were above 25%. This supplies a new benchmark for fixed sounds and an established pair method. It does not establish a common neural cause, semantic coding beyond acoustics, response dimension, single-trial use, or clinical benefit. Clinical coverage, separate cohorts, different sounds, source screens and measurement filters limit scope. Code, permitted electrical inputs, all results, source pins and terms, figures and repeat steps are supplied. Original MRI arrays stay external.

human neuroscience · intracranial EEG · fMRI · sound groups · pair comparison · decision tree · subject holdout · source reuse