What "Golden answer" means

A golden answer is the reviewed, trusted reference response for a query — future runs are graded against it instead of relying on raw string matching alone. It's a 3-step reviewed workflow, not a single click:

1. Capture — save the current run's response as a candidate (not yet trusted).
2. Verify — a human checks the candidate's numbers/facts against the real EHS data and records how they verified it (required — see below).
3. Approve — promote the verified candidate to the approved golden baseline.

Capture and Approve only work when this app is running locally (pipenv run serve) with a writable git checkout — against a deployed environment the write fails safely and this panel will say so plainly, instead of failing silently.

Some queries are graded with a semantic or judge comparison instead of plain text matching — evaluating those makes a real, paid external API call (OpenAI embeddings, or a pinned Claude Haiku judge) every single time that query runs. Rows using either mode are labelled "SEMANTIC" / "JUDGE" next to their golden status.