Every model answers the same frozen question set through the full production agent (find_company / query_database / stats tools against the live graph). A case passes only if the final answer contains the expected ground-truth facts — numbers, dates, names — and none of the forbidden content (raw JSON, tracebacks, hallucinated results). Timeout = no answer within 180 s.

All benchmark questions grouped by category, with each model's verdict per case — click a question to see every model's answer side by side.