Three acoustic model changes failed joint development screens in nine iEEG patients and 34 fMRI listeners

We tested three changes after an earlier spectral model failed independent fMRI confirmation. All records here were development data: nine Podcast iEEG patients and 34 Merlin/Sherlock fMRI listeners. First, four guarded blocks selected ridge penalties within each training record. Second, a declared extension tested three static global-level transforms against spectral and history controls. Third, another extension pooled training-validation losses across people. Every method showed positive clinical gains. None passed the joint fMRI gate for both required contrasts. The two-story fMRI mean spectral gain was -0.00126 with per-record penalties and 0.00068 with pooled penalties. The static-level gain was 0.00204, but its person interval included zero and its matched-control test was inconclusive. The metadata-frozen cohort of 44 new fMRI listeners remained unopened. These exploratory results do not establish cross-modal generalization. The package preserves all failures, input data, ratio grids, validation losses, fitted models, source records, code, and repeat instructions. The scope is selected clinical high-gamma contacts and pooled auditory BOLD in specified offline linear models.

iEEG · fMRI · negative results · model validation · ridge regression · acoustic controls · reproducibility