Policy chatbot (RAG) — synthetic case

Says EUR 30 (policy v3); current v4 says EUR 40

Wrong expense claims, lost trust

Warning banner + fallback for travel questions

Ingestion gap: v4 not indexed (Medium)

Ranking: v3 outranks v4 (Medium)

Model picks v3 from mixed context (Low)

Retrieval trace for the failing question

Contain, then investigate — no rollback

Banner → trace → audit → smallest fix → regression test

Original case passes; zero old-version citations

People Ops + AI platform lead