Design a retrieval benchmark for AI research assistants using known-item search, long-tail evidence, conflicting sources, dates, and paywalled material.

Ordered by provider breadth, then total mentions and owned-domain citations. A mention is not an endorsement or a position claim.

One row per configured provider and category snapshot. Hashes prove answer identity without publishing stored answer text.

Hostnames are derived from the returned public URLs. Best position is the smallest citation position observed.