Thanks!
Hm, you’re getting substantially more divergences than me; I can see you have the non-centered version in your code though. When I re-tool my loop to do single chains as you do and 5000 post-warmup samples, I get:
So, even though they don’t match your data in the overall number of divergences, this latest from me is not nearly as dramatic a difference as I observed previously.
I wonder if it’s related to running chains in parallel vs not somehow? Checking on that now…