can you share the whole model? Or a minimal example that uses LKJCholeskyCov in a similar way?
I am now realizing that in the pieces of code you shared, b0 is not observed, so its posterior samples won’t generally match draws of the distribution we have defined as prior. The prior mean of b0 is 0 for all components yet no posterior sample is close to 0 for any of them and that is exactly what we want!
lkjcc is a hyperprior which always makes things harder to wrap one’s head around, but a prior nonetheless; if you had used a hardcoded prior you would not be trying to match the correlations of the b0 posterior samples to the correlations defined by the prior (our you would be but only to see if they were exceedingly different which could point at prior/data incompatibilities).
Also, sorry for the previous sending on wild-goose chase message with dims and indexing.