Two fixed grammar tests failed in 31 EEG and 27 fMRI listeners; slow MRI terms changed forecast error

We tested whether a context-free grammar part-of-speech score improves within-person neural forecasts beyond trigram and recurrent-network scores and selected speech controls. Separate public Alice studies supplied 31 scalp EEG listeners and 27 fMRI listeners. Six whole-section folds had a 20 s training buffer. Mean added gains were -0.001409 and +47.563 in normalized squared-error units. The 95% person intervals were [-0.002010,-0.000808] and [-16.043,+111.169]. Both fixed shared claims failed: positive section gain and a larger gain with random sample splits. Each Holm p was 1. The MRI baseline itself had very poor section forecasts. A planned ridge check gave opposite gain signs across modalities. A later exploratory removal of slow model terms sharply reduced MRI error, but the added grammar gain remained uncertain. A second regression method reproduced all 58 primary person gains within 5e-11. The package supplies source pins, compact numeric inputs, all results, electrode-template and timing audits, code, figures, and repeat instructions. The tested scope is one narrative and selected prepared signals. No positive cross-modal grammar generalization or causal brain mechanism was confirmed.

human neuroscience · EEG · fMRI · language · part of speech · grammar · cross-validation · reproducibility