OK. Will test. Re. model files, I think that we don't want to distribute a real model file. Instead, the localdev configuration should use really super simple scorer model that does not need to be trained. E.g. one that just takes a single feature and reports it's value. I'll take a look at that.