Brief
Ship a live RAG chatbot over the course corpus: grounded answers with citations, clean OOD refusals, single-agent tool loop, and a public demo recording + runbook.
Dataset
Course corpus index (your m1 benchmark winner config) + your m2 eval set + the 20Q golden suite (must stay green).
Starter guardrails (plan 88)
Winner chunking config only; wrapper + loop skeletons; env secrets; demo script checked in; cost caps on (abort at 5 calls, per-query token ceiling).
100-pt rubric
Live demo works end-to-end 25; groundedness + citations 25; refusals 20; eval report quality 15; runbook + reproducibility 15.
Deterministic checks -> LLM review
Checks: demo URL/recording present; 20Q suite green; eval report has quality + cost columns. LLM review grades answer quality and refusal tone.