Hm, so I ran your slightly modified model (logprior for sigma and logerror) with simulated data and didn’t observe large runtime variations. There was however a correlation between the first two parameters.
Maybe this can become an issue in fitting?
Looking at the chains it seems like they are going to the same place. Given that is only when I have a short run time but in my last few runs they were short all the time. I am not sure what changed. Maybe the issue was my machine since @Funko_Unko has no variation in runtime.
For pair plots it is like @Funko_Unko said, my first two parameters are strongly correlated. I think that was to be expected but I don’t know if that would be a problem in fitting my model.