We have deployed an improved version of the models. The thresholds might need a minor update.
One thing I noted while playing around with the data is that the frequency of edits matching damaging/likelygood is very low for anons (in the single digits monthly, while total anon edits tend to be between 4K-10K). Does that mean the filter threshold is poorly chosen (although it's high for editors, in the 80-90% range), the model is still biased against anons, or does this simply reflect the fact that anons are harder to trust? (It probably doesn't reflect edit quality - manual checks usually find that between a quarter and a third of anon edits are problematic.)
Also, goodfaith/likelybad and goodfaith/verylikelybad are barely different for anons (see graph here showing the fraction of edits these match monthly). They are fairly different for non-anonymous users but then there are (as one would expect) about 100x more matching anon edits. Could this be a threshhold problem, or a bias problem, or is it completely normal?