Image offsets explained 72–89% of early MEG–fMRI covariance in four image sets, but did not pass fixed fMRI tests

We tested whether visual distance maps share structure beyond one additive offset per image. Public Algonauts 2019 targets supplied human MEG and fMRI maps from three experiments, with 15 rows per modality. The first fixed prediction, a shared decrease in the additive fraction from early to later processing, failed. An exploratory covariance split then assigned 71.65%, 86.22%, 89.23% and 75.00% of early cross-modal covariance to image offsets in four image sets. Later shares were 28.50%, 43.06%, 8.24% and 19.99%. These are partitions of observed normalized group maps, not population confidence limits. A new plan preceded access to the third cohort. Five people formed a same-modality template and ten other people tested whether additive covariance exceeded residual covariance. In both twin sets, all ten MEG scores were positive; Holm p was 0.0078125. The fMRI mean scores were -0.00052 and +0.01008, with 95% test-person intervals that included zero and Holm p values 0.96289 and 0.48047. The two twin sets reuse one cohort. The two-modal confirmation rule failed. Image-label controls retained residual alignment, so the result does not show that fusion is false or that additive effects are noise. Correlation-distance bias, source processing, fixed templates, unknown cross-modal row identity and sampled images limit inference. The residual is known U-centering; the new contribution is an empirical audit and inspectable tool. The package preserves code, plans, source hashes, all negative results, figures and repeat steps. External source arrays retain their terms.

MEG · fMRI · representational similarity · U-centering · image offsets · covariance · negative confirmation · visual cortex