In fixed sound sets, full class fits reduced mean prediction error beyond one fitted direction in iEEG and two fMRI studies

We tested the cost of one fitted response direction for five sound classes. Separate public studies supplied signed intracranial EEG means and BOLD fMRI estimates. A local code lock preceded all reserved primary scores. A later audit found that development controls had exposed four reserved electrical patients. We preserve that error and all original results. The audited test uses 14 other patients and 16 MRI adults. For fixed sound sets and source-selected sites, full class fits reduced test loss beyond one fitted direction by 1.146 [0.179, 2.113] and 5.726 [4.820, 6.632] reference-error points. Brackets give pointwise 95% participant Student intervals. The shared gain-and-skill rule passed, with p=0.023726 after Holm correction across the original and audited tests. Eight adults from a second MRI study gave gain 4.939 [4.011, 5.867]. That study shared the MRI sound set. Removing the human vocal class left electrical gain and skill uncertain. Joint person-and-target reweighting gave an electrical gain range from -0.031 to 2.242; it kept fits fixed and does not establish transfer to new sounds. An exact example shows why a prediction gain can occur with a true rank-one test signal. The result does not count neural sources or establish a cause. Clinical sampling, source filters, prior data reuse and full-library site screens limit scope. The package gives code, permitted inputs, source pins and terms, exposure records, all outcomes, figures and repeat steps.

human auditory cortex · intracranial EEG · fMRI · sound categories · reduced-rank prediction · subject-level inference · cohort exposure audit · reproducible benchmark