This page pip-installs the real package and runs it in your browser — no server, no API key, nothing mocked. The same Python that runs in CI is executing here, in WebAssembly, on your machine.
Six questions against a small corpus. Five are answered honestly by the retriever; the last one is a planted hallucination. The harness should catch exactly one. This writes a real eval_run.json — the same artifact the CLI produces — and you can send it straight to the dashboard when it's done.
This is the part worth playing with. Write any answer you like and score it. Faithfulness measures what fraction of your answer's content words are actually grounded in the retrieved context — so a confident, fluent, false answer scores low and gets flagged. Make something up and watch it get caught.