Context
It would be really useful to have predictions for a proposed revision before the revision is saved. One could use this score, for example, in AbuseFilter in combination with other heuristics to decide whether to deny an edit for a certain type of user (0 edits thus far; or creating temp account) or to show a CAPTCHA, etc.
Looking at https://meta.wikimedia.org/wiki/Machine_learning_models/Proposed/Multilingual_revert_risk#Model and https://gitlab.wikimedia.org/repos/research/knowledge_integrity/-/blob/main/knowledge_integrity/mediawiki.py?ref_type=heads, we should be able to provide this metadata in a pre-save context -- we can get the wikitext, a diff against parent revision, the texts that were added/removed/changed, etc.
It would be great to know 1) if Research / Machine-Learning-Team think this is feasible and 2) what the timeline or process might be for implementing something along those lines.
The idea would be to use this with other signals to perform an action. See this document for why preventing edits based on revert risk score alone is problematic.
Related:
- T299436: How impactful would pre-save automoderation be on edit save times?
- T344537: Fast Vandalism Detection
- T123178: [Spike] Investigate building a hook for abuse filter
Proposal
- Add a POST endpoint to Multilingual revert risk and revert-risk language agnostic endpoints to accept metadata that is currently retrieved by querying MediaWiki with a revision ID
Consequences
- Clients can receive revert risk results for an edit before it is saved to MediaWiki