Does it just use the content of the edit itself? Does it consider whether the editor is logged-in or not? The age of the editor's account? The permissions of the account? The time of day the edit was made?
Topic on Talk:ORES/FAQ
@EpochFail@ACraze (WMF) - I bet y'all know the answer to this.
Yes to all of the above. Check out https://ores.wikimedia.org/v3/scores/enwiki/123457/damaging?features
You can experiment with the local gradients for each feature by injecting counterfactuals: ORES/Feature injection
For a more systematic way of exploring how the features lead to a given prediction, there are libraries which vary each feature and show its impact on the output. For example: https://github.com/adamwight/ores-lime/blob/master/Explain%20edit%20quality.ipynb