References from sites with weak training contrast failed a shared prediction test in six ECoG patients and six fMRI participants

We tested whether weak category sites can supply a useful reference for a strong-site category reader. Public studies supplied separate human ECoG and fMRI cohorts. One ECoG patient supplied development. The fixed test used six other ECoG patients and six fMRI participants. The reference used the weakest half of sites and had zero face-house mean contrast in training. Its fitted coupling to strong-site category residuals set a half-strength correction. Mean relative test loss decreases were 19.10% (2.03% to 39.32%) in ECoG and 16.61% (7.05% to 25.93%) in fMRI. Parentheses give pointwise 95% person bootstrap intervals. Gains over expected within-category reference reassignment were 24.76% (3.17% to 51.67%) and 23.35% (7.16% to 41.93%). The single shared claim required both gains in each modality. Its exact person sign-test conjunction p was 0.34375; the rule failed. A zero training contrast does not establish absent category information in later data. Clinical sampling, small cohorts, repeated images, selected fMRI masks, different response durations, and offline processing limit scope. This is a fixed new-endpoint test in reused cohorts, not new-study replication or evidence for a common causal source. The package has code, source hashes and terms, all person results, controls, figures, checks, and repeat steps.

ECoG · fMRI · category decoding · weak sites · correlated variation · reference correction