Fixed GPT-2 tests did not confirm shared order gains at selected sites in five ECoG patients and ten fMRI adults

We tested whether intact past word order adds neural prediction value. Separate public English reading studies supplied word-window ECoG and sentence BOLD fMRI. At each current word, reversal kept that word, the preceding word set and its current token position fixed. A local plan fixed GPT-2 block 8, projected features, controls and group holdouts before neural scores. Mean order gains were -0.0407% and +0.1508% of null loss in five ECoG patients and ten fMRI adults. Their 95% person t intervals were [-0.2395, +0.1581]% and [-0.0370, +0.3386]%. The fixed shared order-and-context rule failed, p=0.96875000. Mean intact-model held R2 was 0.0075 and -0.0297. Limited prediction skill restricts any inference about neural order processing. Removing position and length terms raised the ECoG order mean to +0.3548% in a fixed descriptive check. All negative scores and fixed checks remain. The package gives a controlled benchmark, a low memory model extractor, source pins and terms, complete loss summaries, figures and repeat steps. Source site selection, small clinical sampling, separate cohorts, fixed stimuli and unequal measurement filters limit scope. The model manipulation was not shown to people. It cannot establish a common neural cause. Original neural arrays and model weights stay external.

language · word order · ECoG · fMRI · encoding models · GPT-2 · independent studies · negative results