Thanks @bwengals and @cmgoold! I’m a bit more experienced now with the general idea of the Dirichlet but this is good to see!
In a recent post here @tcapretto indicated PyMC v5 has fixed and updated the ordered transform for pm.Normal. I’ve been a bit busy working on that project but the code I imagine is the same; I wonder now if v5 would work with a Normal prior on the cutpoints now?
I guess the issue is that there is a need to fix the top and bottom cutpoints to constants, and the question is what (necessarily) multivariate prior sits between those constants and how this affects sampling/inferences! I’ll try to find some time to explore this and report back.