At a 5% site loss budget rounded up, AUC bounds exceeded 0.5 in three of six ECoG patients and all six fMRI adults

We tested whether a frozen face-house reader keeps ranking above chance after any fixed set of sites is lost. Separate public studies supplied human ECoG and fMRI. One ECoG patient supplied integration checks. A local code lock preceded new loss endpoints in six other patients and six MRI adults. The primary loss count was 5% of sites, rounded up. A conservative all-set AUC bound exceeded 0.5 in 3/6 patients and 6/6 MRI adults. The fixed shared sign test failed, p=0.65625. Mean bounds were 0.5076 and 0.9954. The pointwise 95% person bootstrap ranges were [0.4759,0.5432] and [0.9861,1.0000]. An exploratory exhaustive search found true worst test AUC above 0.5 in five ECoG patients. One patient fell from 0.561 to 0.488; random losses averaged 0.561. All losses of one site kept test AUC above 0.5 in all 12 final people. This is a finite-record audit, not independent confirmation of a population effect. Exact examples show that the bound can be loose and that copies of one scalar can give perfect site-loss robustness. The contribution is a checked bound tool and a neural readout benchmark. No positive shared 5% certificate, independent neural redundancy, or cause was established. Small selected cohorts, different tasks and measurement units, source masks, repeated images and missing artifact records limit scope. Code, source pins and terms, all results, figures and repeat steps are supplied. Source arrays stay external.

human neuroscience · ECoG · fMRI · visual category · site loss · robustness bound · negative result · reproducible benchmark