A 1 s word history term raised mean test error under fixed models in six iEEG patients and 25 fMRI listeners

We tested whether recent word history improves neural prediction during story listening. Three iEEG patients selected a 1 s history term. We fixed the term before new fits in six other iEEG patients and 25 fMRI listeners from another study. Each person had separate ridge models. Controls included sound bands, amplitude shape and history, word timing, lexical variables, and 64 common-word identities. Mean test R-squared change was -0.000200 in iEEG (95% person interval -0.000315 to -0.000114) and -0.001789 in fMRI (-0.003183 to -0.000356). All six electrical gains were negative. Eight of 25 BOLD gains were positive. Both planned positive-gain tests failed. The result limits this fixed reader, feature, sampled sites, and two stories. It does not establish absent neural memory or a causal modality effect. Five fixed controls per modality and every negative score are retained. The package supplies an inspectable benchmark, pinned source records, numeric inputs, code, plans, figures, and repeat steps. All 75 external fMRI derivative files matched source hashes. A separate direct fit repeated all 31 primary person scores.

word history · intracranial EEG · fMRI · natural speech · negative result · prediction · independent confirmation · reproducible benchmark