A fixed rule from six ECoG patients gave higher mean phrase prediction error than linear and log rules in 55 fMRI adults

We tested whether a mean electrical response over successive words predicts the fMRI response to phrase length. Separate public studies supplied six ECoG patients and 55 MRI adults. A local plan and code lock preceded new numerical results. The electrical model used the sentence-minus-word-list response and a reset-and-average assumption. MRI lengths 1 and 6 set each person's response level and scale. Lengths 2 and 4, and length 3 where present, were held from calibration. Three MRI panels contained 15, 25 and 15 people. Mean electrical-rule losses were 0.33968, 0.60331 and 0.35621 squared BOLD effect units. Linear-rule losses were 0.17450, 0.27104 and 0.21239. Log-rule losses were 0.12653, 0.19409 and 0.21375. The fixed compound improvement test failed, p=0.99990845. Electrical context growth was positive in all six patients, with mean 0.13086 and 95% person limits [0.05642,0.20529]. Thus known context effects did not validate this quantitative link. Source-person uncertainty was wide. One electrical trial half had an unstable scale. A fixed all-valid-electrode control improved the losses but could not replace the main test. A later MRI region-block check retained a reversed new-batch comparison against the log rule. All negative and discordant results remain. The contribution is a checked prediction tool and a complete negative benchmark. Separate tasks and cohorts, clinical spatial coverage, source electrode selection, omitted electrical response tails, a table-label assumption and measurement filters limit scope. No positive neural generalization or causal result was confirmed. Code, source pins and terms, derived scores, uncertainty draws, figures and repeat steps are supplied. Original neural inputs remain external.

human neuroscience · ECoG · fMRI · language · sentence context · phrase length · cross-study prediction · negative result