Comment by quinncom

AI;DR

> That distinction matters when you're debugging an agent, because "the answers got worse after the model swap" is not one failure, it's five