Fixed action and object label products failed a shared test of higher error in five EEG and eight fMRI adults for 102 videos

We tested whether a product of action and object labels changes prediction of human brain responses to videos. Two public studies supplied separate EEG and BOLD fMRI cohorts. One EEG adult and two MRI adults supplied development. All three had higher error with the added term, so a local code lock fixed that direction before tests in five other EEG adults and eight other MRI adults. Models fit up to 1000 clips and predicted 102 other clips. A fixed comparison model included separate object, action and scene terms, their quadratic terms and two scene products. At a unit component weight, added mean error was 0.07118 and 0.06045 percentage points of a fitting-mean reference error. Their 95% person Student limits were [-0.03840, 0.18075] and [0.00686, 0.11404]. The shared gate p was 0.1456494. The fixed product term did not pass the shared error-increase test. This test used a fixed panel of 102 clips. A check that sampled people and test clips gave ranges that crossed zero in both studies. The EEG scope was 35 posterior channels at 150-350 ms; MRI used bilateral STS. EEG clips had audio; MRI clips were silent. All specified checks and discordant people remain. A fixed synthetic product source gave a 3.10-point gain. An exact confounded-feature model shows that a gain need not identify neural binding. The contribution is a compact benchmark and a tested model comparison. Fixed priors, source processing, sensory differences, response precision and small cohorts limit scope. The result does not establish absent binding or causality. Code, compact neural means, source pins, results and repeat instructions are supplied.

EEG · fMRI · visual events · encoding models · action and object labels · independent confirmation · public data