Dependence of convergence on random seed

[junpenglao]

@adamh

My current understanding is that there are “multiple possible solutions”, and the bad mixing is an indication of the sampler finding different solutions from different starting points - is this a plausible explanation?

Likely that’s the case.
Do you have normalization in your model like x/x.sum()? if so you should change them into better parameterization.
Order constraints are better done using transform.Order