Fixed tests did not confirm four-axis finger gains in four ECoG patients and six fMRI adults

We tested whether four fitted axes reduce later finger-response error compared with two axes. Separate public studies supplied ECoG and 7 T fMRI. Two patients and one MRI adult supplied development. A local code lock preceded reserved endpoint tests. Fixed count rules left four ECoG patients and six MRI adults. Mean gains were +0.2029% and -0.1115% of baseline error. Their 95% person t intervals were [-0.2828,+0.6886]% and [-0.2214,-0.0016]%. The fixed shared test failed, p=0.984375. Four electrical patients also made its exact sign-flip threshold unattainable at 0.05. We retain that design limit. An added exploratory audit kept the two-axis gain and fitted a separate gain for extra axes. It changed the MRI mean comparison to +0.0138%, with pointwise interval [+0.0036,+0.0241]%. Electrical uncertainty still included zero. No positive cross-modal generalization was confirmed. Two exact models limit a rank interpretation. Specified block errors gave an 80% model gain from a true two-axis signal. A true one-axis signal gave a 4/65 gain with noiseless calibration and test signals. All eight signs of independent training errors gave this result. The training errors had zero mean. These models were not fitted to people. Separate cohorts, clinical sampling, unequal task and measurement scales, source processing, cue labels and small samples limit scope. The contribution is a reproducible benchmark, a calibration audit and exact interpretation checks. The package preserves code, source pins and terms, plans, all outcomes, figures and repeat steps. Neural arrays remain external.

human neuroscience · ECoG · fMRI · finger responses · dimension reduction · calibration · cross-validation · negative result