As an abusefilter maintainer, I would like to be use the likelihood of a revision being damaging and/or goodfaith to decide whether a filter should display a warning, disallow the edit, or do nothing.
This would allow filters to have less false positives than manually constructed rules which take into account only a small subset of features which revscoring models can use.
That would be cool!
This is technically possible, yes. But I'm skeptical, because AIUI ORES is currently set up to process edits *after* they're saved, how would it work pre-save? Use PST wikitext? HTML? What about content that is vary-revision?
Also we're working on cutting down save time a lot, and adding an extra network request may work counter to that goal...
I would like this to be reconsidered, as I don't see the problem: there are many features¹ in revscoring which seems to be available to AbuseFilter already as variables, or which could be obtained from wherever AF gets them right now.
At most we would need to train specific models not requiring any feature for which the date is only available after saving the page. Anyway, this would provide AbuseFilter a model which is a lot better than our current strategies of checking *only* the namespace and the number of bytes in each revision:
https://pt.wikipedia.org/wiki/Special:AbuseFilter/3
https://pt.wikipedia.org/wiki/Special:AbuseFilter/119
¹ "feature" in this sense: https://github.com/wiki-ai/revscoring/blob/master/revscoring/features/wikitext/features/chars.py