Eval as an Input, Not a Dashboard: Building Self-Healing LLM Systems
Most teams treat evaluation as a scoreboard — a number you glance at and feel good or bad about. The frontier idea is to wire the eval back into the system as an input that rewrites it: traces go to independent judges, fixes get proposed automatically, and one human approves. It's AI all the way down, with a single gate that isn't.

