We are increasingly seeing variations in errors in regression accuracy tests due to very small - mathematically equivalent - changes in the model, but worse, between Julia versions. A more detailed discussion is given in #300 . In the long term this has to be addressed, but it is currently not clear how.