Student tests found mean motor ranking above chance after feature removal in 13 ECoG patients and 58 fMRI adults

We tested hand-versus-tongue ranking after the half of measured sites with the largest training contrasts was removed. Separate public studies supplied scalp-referenced 76–100 Hz ECoG power and BOLD fMRI maps. Three patients and four MRI adults supplied development. A local code lock preceded the new endpoint in 13 other eligible patients and 58 other adults. Two groups selected features and fitted a linear reader; four other groups tested it. All 15 splits gave one score per person. Weak-panel mean ranks were 68.846 [62.485, 75.208]% and 96.695 [93.678, 99.713]%. Brackets give pointwise 95% person Student limits. The shared above-chance mean test passed, three-claim Holm p=9.41633e-05. The original half-performance retention rule failed, adjusted p=1; the joint majority rule failed, adjusted p=1. All failures remain. These are ranks of averaged response maps, not single-trial accuracy. Independent scalar loops checked primary arithmetic. A one-source spatial-blur model passed despite exclusion of the latent source site. Thus measured feature tolerance does not establish redundant neural sources or lesion tolerance. Small clinical samples, unequal movements and coverage, estimated feature ranks, shared reference, source smoothing and stable artifacts limit scope. The package supplies a checked two-study benchmark, code, complete scalar results, source pins and terms, figures and repeat steps. Original neural arrays remain external.

human motor cortex · ECoG · fMRI · feature removal · held-out ranking · reproducibility