Our key assumption? We imagine that evaluation questions are randomly drawn from an underlying distribution of questions. This assumption unlocks a rich theoretical landscape, from which we derive five core recommendations.

→ View original post on X — @anthropicai