Evaluation and feedback

Two loops tell us whether Flo is getting better:

Loop	When	Question it answers	Results
Eval on every merge	After each staging sync	Did this change break routing or grounding?	Langfuse evaluation project, Experiments
Feedback and triage	Every user 👍/👎, triaged nightly	Where do real answers fail, and who should fix it?	Langfuse live project, scores and a report

All judging uses flo-judge (Qwen2.5-7B), a different model from the one that answers, on its own GPU. See Models.