A fixed 32-sound selection rule failed its shared accuracy test in 21 iEEG patients and 16 fMRI adults

We tested whether acoustic features can choose 32 calibration sounds that improve neural response forecasts. The design combined variance criteria from four penalties. Each panel chose its own fit penalty from calibration sounds only. Separate public human studies supplied signed iEEG LFP means and BOLD fMRI. Four people per study supplied development. A local code lock preceded tests in 21 other patients and 16 other adults. We removed broad sound class means at scoring. Mean error gains over random panels with the same sound class counts were 1.123 and 0.988 points of observed contrast energy. Their 95% person Student intervals were [+0.239, +2.008] and [+0.660, +1.316]. Mean chosen forecast skills were 2.542 and 5.711 points; the electrical interval included zero. The fixed shared rule for positive gain and skill failed, Holm p=0.269852. A fixed rule for a mean gain below one point also failed. A bootstrap over people gave a positive electrical skill range, but person and sound resampling included zero. The MRI gain range also included zero under that resampling. All negative people and planned checks remain. Direct fits and fresh source chunks checked the numerical results. The contribution is a new fixed benchmark and a reusable implementation of standard design ideas. It does not establish shared useful prediction, absent acoustic coding, a common neural cause or savings in recording time. Reused cohorts, source screens, unequal sites, sound corpora, measurement filters and conditional uncertainty limit scope. Code, permitted inputs, source pins and terms, all outcomes, figures and repeat instructions are supplied. Original MRI arrays and audio remain external.

human neuroscience · intracranial EEG · fMRI · stimulus selection · acoustic prediction · regularization · negative result · reproducibility