With no bounds on word effects, a word model fits all sentence means in one ECoG and one fMRI reading dataset

We tested whether the rarest word improves neural prediction beyond additive lexical frequency measures. Separate public reading studies supplied ECoG and fMRI. A local code lock preceded new scores in five ECoG patients and eight reserved MRI participants. Mean gains were -2.234% and -0.171% of null loss. Their 95% person Student intervals were [-4.215,-0.254]% and [-0.490,+0.148]%. The shared positive rule failed, p=1. Baseline prediction was weak. An exact design audit found a different shared result. The word-count matrices had full row rank: 52 of 52 ECoG sentences, 243 of 243 and 384 of 384 MRI sentences, and 627 of 627 in the MRI union. Hence every sentence response vector, including noise, has an additive word fit if word effects have no bounds and can differ between neural sites. Any test response vector can coexist with fixed training responses. Nonnegative effects do not restore a test if the intercept is free. Exact modular certificates and fits to all finite site records checked these statements. Minimum coefficient bounds were 2.967-6.056 and 3.356-5.056 network standard deviations under a stated scale. The original ECoG word rows retain 153 residual dimensions after word identity, position and source-order control; sentence averaging leaves none. A new four-sentence design cancels all fixed word effects but permits a maximum-rule contrast in a model example. No participant saw that design. This is a method contribution and a negative cross-modal benchmark, not evidence for an additive brain code, absent composition, or causality. Code, source pins and terms, all results, exact certificates, figures and repeat steps are supplied. Source arrays and frequency tables stay external.

human neuroscience · ECoG · fMRI · language · word effects · identifiability · sentence averaging · cross-modal tests · negative result