A fixed square root word surprise term failed a shared gain test in seven iEEG patients and 18 fMRI adults

We tested whether a square root of model word surprise improves response prediction beyond linear surprise, sound and word controls. Separate public studies supplied intracranial high-gamma amplitude and BOLD fMRI during story listening. One fixed GPT-2 model supplied new probabilities for both stories. Two patients and four MRI adults supplied development. A local code lock preceded tests in seven other patients and 18 other adults. Mean added gains were +0.011794 and +1.498866 percentage points of held null prediction error. Their 95% person Student limits were [-0.013599,+0.037186] and [-0.928712,+3.926444]. The shared rule failed after correction for two claims, p=0.328125. A linear gain test also failed, p=0.984047. MRI baseline prediction was weak, and one known 1% curve injection failed in development. A constructed noiseless example gave negative gain when the added coordinate was the true response. Thus the failed reader does not show that brain responses are linear. The package gives new word estimates, audited inference with memory-mapped weights, source pins, every person score, negative results, figures and repeat code. Clinical coverage, separate cohorts and stories, source processing and short MRI records limit scope. No positive neural shape generalization across modalities or common cause was established.

word surprise · surprisal · iEEG · fMRI · language · neural encoding · model validation · reproducibility