"""FinanceBench benchmark: gold questions, answers, and evidence in Postgres.

`load` and `pages` fill the `financebench` schema, which retrieval and the
answer agents never touch. No code in the package reads its questions or
answers back. The answering path is measured instead against a gold set that
`verify_gold` builds from agent-authored questions checked against the cells
of ingested documents, and `run_shapes` scores the answer shapes against it.
"""