A fixed test did not confirm a 1% error limit for compact sound models in 18 iEEG patients and 16 fMRI adults

We tested whether at most four learned acoustic response patterns preserve prediction of responses to sounds excluded from fitting. Separate public studies supplied signed intracranial EEG local field potentials and fMRI BOLD estimates. Four people per modality supplied development. A fixed test used 18 other patients and 16 other adults. Nested sound folds selected rank and ridge penalties. The full rank comparison chose a separate penalty for each site. Mean extra compact error was 1.005% and 0.915% of intercept test error. The 95% person Student intervals were 0.425% to 1.585% and 0.557% to 1.274%. Neither modality established a mean cost below the fixed 1% margin; shared p was 0.50669. Compact models retained positive mean prediction skill. Spatial mean removal reduced the electrical cost. Five-site BOLD panels and separate BOLD estimates changed the mean ranking. These controls reuse people and do not repair the failed primary test. The work supplies a new benchmark, permitted electrical inputs, all person and sound errors, source pins and terms, figures, code and repeat instructions. The estimator is established. Cohort, feature, coverage, signal and source processing differences limit scope. This result does not establish a shared neural dimension, a universal error limit, or causality.

human neuroscience · intracranial EEG · fMRI · auditory cortex · reduced rank regression · acoustic encoding · non-inferiority · subject-level inference · cross-validation · negative result · reproducible benchmark