Position: Claims, Not Papers, Should Be the Computational Unit of the Scientific Record

Position. Scientific knowledge should be represented at the level of claims and their evolving relations, rather than treating papers as the canonical computational units of the scientific record. Papers are valuable communication objects. They bundle motivation, methods, evidence, interpretation, and argument into narratives optimised for human readers. But the scientific standing of a paper's claims evolves separately. A central claim may be independently replicated, a subsidiary claim contested, a methodological assumption superseded, and another conclusion simply repeated without new evidence. Paper-level representations collapse these trajectories. This mismatch predates artificial intelligence, but AI scientists make it urgent. Granular scholarly representation has been attempted for fifteen years, from citation typing to nanopublications to knowledge graphs, and has not displaced the document. We argue the failure was one of cost and demand rather than of representation: annotation was expensive and almost nothing consumed the output. Both conditions have now changed. We argue that the computational substrate of AI scientists should be a living claim network in which support, contradiction, replication, refinement, dependence, and supersession are first-class relations, with provenance and evidence retained at claim level. We make this concrete with one recommendation aimed at this community: benchmarks for AI scientists should score change in the knowledge state, measured against a claim network, rather than paper generation or peer-review acceptance. We give an operational definition and a retrospective held-out protocol, and we take seriously three objections the position must survive: the Duhem-Quine problem for contradiction edges, the interpretative character of independence judgements, and the fact that a claim network is a cheaper target for gaming than a citation count. The proposal is not to abolish papers, which remain the human-facing narrative interface. The underlying machine-actionable scientific record, however, should become claim-centric.