And thanks to my coauthors Fred, Yi-Shyuan, Hari, @ceshwar.bsky.social , and @kkarahal.bsky.social for helping to make this paper happen

I think the broader framework for human-AI collaboration where we use predictive models to surface misalignments between decision-makers is a super promising one. Come talk to me about it!

One unexpected use-case for the system that came up in qualitative interviews: on-boarding new moderators.

Using simulations on static datasets, we show that under ideal conditions, Venire can approximate the decision consistency benefits of universal panel review while assigning only ~30% of cases to panel. Under less ideal conditions, this number increases to ~60%.

The paper itself is a kind of a three part journey. We discuss initial interviews with mods where we surface attitudes towards decision consistency as a policy outcome, plus quantitative evaluations of Venire's ML model, and a think-aloud study where moderators used Venire in a sandbox modqueue.

Moderation teams are usually resource constrained, so case-level rule enforcement is often handled by individuals. Venire's core goal is to do effective triaging -- identify the subset of cases where panel review is most urgently needed.