A fixed offset update did not pass a shared heat forecast test in 115 EEG and 20 fMRI adults

We tested whether an unlabeled late response batch can lower heat forecast error on other rows. Separate public studies supplied scalp EEG and BOLD fMRI. Spatial readers fit early records from 39 EEG and eight MRI adults. A local code lock preceded new endpoint tests in 115 other EEG and 20 other MRI adults. The fixed rule required positive mean bounded offset gain and updated forecast skill in both studies. It failed, p=0.31891. Mean gains were 0.01031 and 0.00445, with 95% person Student limits [0.00056,0.02007] and [-0.00897,0.01787]. Training-and-target person resample ranges crossed zero in both. A fixed exploratory sparse-reader test in 38 records from another MRI study gave -0.00127 [-0.00774,0.00519]. All negative people and failed checks remain. A post-analysis quadratic risk tool covers every common update fraction from zero to one. Under four simultaneous person Student limits, its upper mean gains were 7.91% and 8.33% of zero-forecast risk. These bounds concern the fixed reader class, not all neural information. Checked model examples show that a shift can fail to help and that shared nuisance can produce gain without neural drift. This is a negative cross-modal forecast benchmark and a reproducible method tool. No positive shared neural effect or cause was established. Reused cohorts, source selection, full-record preprocessing, short calibration batches, trial dependence, missing motion traces and different measurement filters limit scope. The package supplies licensed compact inputs, source pins and terms, code, complete results, figures and repeat instructions.

human neuroscience · EEG · fMRI · heat forecast · unlabeled calibration · dataset shift · negative result · reproducibility