Mean gains from fixed quadratic terms were below one error point in six SEEG patients and 16 fMRI adults

We tested whether quadratic neural features improve prediction of sound responses. The baseline used four linear response coordinates. Inputs and outputs used different sites and repeats. Separate public studies supplied human SEEG and BOLD fMRI. Development used two patients and four MRI adults. A local plan and code lock preceded the reserved tests. Each person supplied 12 input sites and 12 output sites. Mean gains were +0.0128 [-0.2159, +0.2414] points in six reserved patients. They were -0.0489 [-0.1380, +0.0401] in 16 reserved MRI adults. Brackets give 95% person Student limits. One point is one percent of the specified category-only squared-error baseline. The positive shared claim failed, Holm p=0.870895. After development, we declared a one-point mean-gain ceiling. This ceiling passed in both modalities, Holm p=0.03125. The tests assume person independence and null sign symmetry. Eight other MRI adults gave +0.0589 [-0.0693, +0.1870]. A separate exact model gives a 5.3519-point quadratic gain with linear latent signals and independent input noise. It is not a fit to human data or the full finite benchmark. All negative people, controls, and a failed sound-identity join remain. The finding limits one regularized readout. It does not establish linear brain computation or a common cause. The package preserves code, source versions and terms, losses, uncertainty, figures, audits, and repeat steps. Original neural arrays remain external.

human neuroscience · SEEG · fMRI · sound responses · quadratic prediction · measurement error · cross-validation · reproducibility