At a stated 0.10 error limit, a shared sound area test failed; cyclic differences gave positive SEEG and fMRI controls

We derived a sharp error bound for a prior signed response area test. It subtracts eta/[4(1-eta)] times repeat disagreement. The bound needs a common response mean, zero conditional error means after training and selection, and two bounded cross-error moments. It allows unequal variances and non-Gaussian errors. Separate public studies supplied eight SEEG patients and two fMRI groups. The MRI groups had 20 and eight adults. A local code lock preceded new scores. At eta=0.10, mean margins were +0.043791, +0.257556 and +0.091270. Their 95% person Student intervals were [-0.015121,0.102703], [0.188065,0.327048] and [0.042209,0.140331]. The fixed shared test failed; two-claim Holm p was 0.113720. The unadjusted shared test also failed. Positive controls did not replace these failures. An added exploratory test used cyclic differences of three repeats. All 36 resulting margins were positive, with means +0.204276, +0.160672 and +0.089573. An exact rank-zero Gaussian model explains why this control can fail: the two test differences sum to the negative training difference. Its expected adjusted margin was 0.143333 in a stated example. The new contribution is the sharp bound, its sensitivity tool, a negative human benchmark and a checked difference-control failure. Source screens can violate the required conditional means. Thus no shared neural rank or cause is established. Code, source pins and terms, all outcomes, figures, plans, audits and repeat instructions are supplied. Original neural arrays remain external. Student tests can have incorrect rejection rates. A skewed rank-zero model rejected 7.2% of 10000 eight-person null groups at a nominal 5% level.

human neuroscience · SEEG · fMRI · response dimensions · repeat errors · sensitivity analysis · negative result · cyclic differences