M1–M6 failure taxonomy · LLM-as-judge eval gate in CI
Pickle Playbook
RAG strategy coach for pickleball players, deployed on Render/Vercel/Supabase. Failure-mode taxonomy (M1–M6) + LLM-as-judge eval gate in CI that [INSERT EVAL NUMBER — e.g. caught X% retrieval failures / lifted answer quality Y→Z on frozen eval set / blocked N regressions].