A fixed name prior reduced error with one class removed from each fit in 14 iEEG patients and six FFA-1 fMRI adults; electrical total skill was uncertain

We tested whether object-name similarity helps predict a class omitted from each person's response fit. Separate public studies supplied human depth EEG and fMRI. Development used three patients and two MRI adults. Classes could occur in development. A local code and settings lock preceded tests in 14 other patients and six other adults. A ridge prior from a fixed text model was compared with unrelated labels. Mean added gains were 0.066% and 1.648% of training-mean error. Their 95% person intervals were [0.014, 0.118]% and [0.961, 2.334]%. The shared exact sign-flip test passed, p=0.015625. Electrical total skill was only 0.047%, with an interval that included zero. Its class/person sensitivity range also included zero. A pre-image electrical gain had a larger point estimate but wide uncertainty. Random name assignments and all planned controls remain. A later exact calculation showed a positive comparative gain under independent noise in the electrical label designs and a negative gain in the MRI design. This calculation is not a human bias estimate. A later matched-penalty electrical check included zero. Thus, the evidence supports a comparative rule in fixed processed panels, not a shared semantic code or a cause. Clinical VTC and selected FFA-1, separate cohorts, different images, source selection and measurement filters limit scope. The package gives all results, source pins and terms, code, figures, checks and repeat instructions. Original neural arrays and model weights remain external.

human iEEG · fMRI · class exclusion · object names · encoding models · prediction variance · independent test · reproducible research