Today, we are launching IRIS, our open-source paper-finding agent. Ask IRIS a literature question in plain English, with all of its conditions attached. It returns a ranked list of papers that were checked against every condition, one by one, with the sentences it relied on quoted next to each result. Keyword search cannot do this. It returns papers that happen to use your words, misses the ones that say the same thing in other words, and has no way to check a condition like "evaluated on CASP targets after 2022," so you end up opening papers one at a time to check for yourself. IRIS sends each question down the path built for it. Venue, year, author and citation conditions become set arithmetic over the corpus and are executed exactly. Topical questions get dozens of separate searches, one of which writes a hypothetical abstract of the paper that would answer the question and searches with that instead, and every candidate the searches return is scored against each condition on its own. We ran IRIS on PaperFindingBench, the paper-finding task in AstaBench. The benchmark asks 66 literature questions of three kinds, each with a list of the papers that answer it, and scores a system on how many of the right papers come back and how highly they rank. A language model grades any returned paper the list does not cover. IRIS scores 0.382 on the validation split at $0.13 a query, the strongest result published on that split. On questions naming one specific paper it scores 88.00, and on venue, year, author and citation questions it scores 68.24. Topical questions with several conditions attached come in at 22.78, and the reason is retrieval, since a paper that never enters the candidate pool cannot be ranked out of it. About 30 percent of the correct papers are never returned by any search built from the question's own words. Everything we benchmarked is the code we published, under an MIT license, so you can clone IRIS and run that configuration yourself today.