Exactly :) Proper (in this case, extensive) distribution sampling and OOD evaluation are essential to ensure that brain-model correlations capture relevant abstractions. Local statistics go a "surprisingly" long way.