A shared word boundary prediction failed in seven iEEG patients and 18 fMRI adults under fixed response filters

We tested whether word surprise predicts neural responses better at a word's acoustic end than at its onset. Separate public studies supplied human high-gamma iEEG and BOLD fMRI during two stories. A local plan and code lock preceded tests in seven patients and 18 MRI adults. Both models used the same sound and word controls. Mean end advantages were -0.125904 and -0.008713 points of held null loss. Their 95% person Student limits were [-0.270727,0.018918] and [-0.024729,0.007303]. Both shared direction rules failed, Holm p=0.97656250 for end and 0.63153076 for onset. MRI prediction was weak. Known 1% signal additions gave wrong timing preferences in some development MRI records. An input-only tool gave squared canonical correlations of 0.997070 and 0.949505 between the MRI timing spaces under fixed controls and a reference clock. An exact midpoint made the most similar unit pair ambiguous at bounded projected error 0.027074. This is not an estimate of MRI noise or a universal timing limit. A duration-weighted derivative removed 94.1% of the Alice input boundary difference under a smooth BOLD filter. A second duration moment removed 99.6%. A hypothetical BOLD filter on the electrical story gave the same type of design result. No new Podcast MRI data were added. Independent arithmetic, a two-person fresh source rebuild and a clean compact repeat checked the work. The package supplies code, source pins and terms, compact inputs, all negative results, figures and repeat steps. No positive neural timing generalization or cause was established. Separate cohorts and stories, clinical sampling, full-record preprocessing, broad pooling, uncertain source clocks and fixed response filters limit scope.

human neuroscience · word timing · iEEG · fMRI · surprisal · model ambiguity · negative result