(mind you, this is already log likelihood, but the summing pattern across datapoints when far from the maximum gets pretty nasty)

as well as this, allowing the nifty fp192* format

*: conditions apply

i think it's fine, application-specific pre/post-processing are convenient; and for more complex models, it could very well be faster for prototyping ("real" programming language vs DSL where you forcefit what you want to do in an existing construct). might drop a Stan version when finished, though

you know you're overperforming when the audience is pulling up your google scholar /during/ the talk #Hünfeld2025 (session 1)

(Joke aside, constant pH MD and protonation coupling talk tomorrow, excited!)

And a documentation website, too

grubmueller.pages.mpcdf.de/docs-gromacs...