One wrong example left uncertain mean calibration gain in 20 iEEG patients and positive gain in 16 fMRI adults

We tested whether one wrong-class sound preserves a four-example response correction. Separate public studies supplied signed intracranial field means and BOLD fMRI estimates. A local code lock preceded new scores in 20 iEEG patients and 16 fMRI adults. Wrong examples came from acoustically matched training candidates. Correct support draws were averaged exactly. Clean mean gains were 4.675 and 10.367 error points. Random replacement gave 1.171 points in iEEG, with 95% person Student limits [-3.773, 6.115], and 9.014 points in fMRI, with limits [8.429, 9.599]. The shared retention rule failed, Holm p=0.778042. An oracle selected one wrong candidate by mean test loss. Its iEEG gain was -20.102 points, with limits [-59.990, 19.786]. All 16 MRI adults retained positive oracle gains; their mean was 6.663 points. The second shared rule also failed. An added eight-adult MRI study retained positive mean gains. One electrical patient had random gain -40.772 points. A later training-only norm choice gave a similar large error. In six affected settings, one wrong sound had 794 to 6767 times typical correct-support residual energy. This is large leverage in processed measurements, not an identified artifact or neural cause. Cohorts, class panels, acoustics, source screens, temporal filters and site coverage differ. Fits stay fixed in sound-resampling checks. No positive cross-modal tolerance claim was confirmed. A later coordinate-median correction gave random-replacement mean gains of 3.587, 9.258, and 9.580 points in the same iEEG and two MRI samples. The electrical worst-candidate mean remained uncertain. This median comparison is exploratory and has no new independent neural confirmation. The package supplies exact stress tools, licensed electrical summaries, source pins, all negative results, code, figures and repeat steps.

human neuroscience · intracranial EEG · fMRI · auditory encoding · calibration · label errors · robustness · finite support