Here is a screenshot of the one time the optimization worked:
AHA! I just figured out that the Gamma distribution is parameterized as Shape, Scale and not Shape, Rate (as in R)… so fixing that may solve my problem since the prior for the size of the errors was parameterized by a gamma which was supposed to have small scale…
Also discovered some plain old typos/bugs. Looking forward to seeing what happened with the sampling run from last night…
Ok, it seems that having fixed some bugs at least the fits aren’t ridiculous after optimization, and it’s reliably a reasonable starting spot:
I’ll open a different question related to how to usefully work with chains with tens to hundreds of parameters.