coelacanth
independent. sep 2026 - present.
a benchmark for systems that check whether a scientific idea has been done before. it pairs claims from one field with earlier work from other fields that may describe the same thing under another name. systems see only the claim, in its own field’s vocabulary.
every pair carries a typed relation, from equivalent to nothing close, backed by quotes that code checks verbatim against the source. people choose every label. each item is also put to language models with no search, to tell retrieval from recall.
built to measure how often existing novelty and precedent checkers miss cross-field rediscoveries.
the paper, pdf ↗18 pages