I argue that we need to account for the size of the model space when determining sample size, as larger model spaces reduce power. I also show that the commonly used “fixed effects” model selection approach is statistically unreliable. An analysis of the literature suggests shortcomings in both

More generally, we link MEC coding to planning-ready compositional representations, with invariant and modular responses in ubiquitous MEC object vector cells. These cells provide the building blocks of compositionality in the model.

Neurally, influential work proposed grid cells encode eigenvectors of the successor map. Nice idea, but it struggles when barriers or goals change. Our model ties grid code to the compositional map, keeping them useful even as the world changes, consistent with empirical findings on local remapping.