The claim graph as a first-class artifact

Abstract

The paper-as-atom convention served citation but is the wrong granularity for the queries readers and agents now run: has this specific result been replicated?, what does the literature say about this sub-question?, which downstream work depends on this contested step? We argue that scholarly knowledge should be addressed at the claim level, with each registered assertion a first-class node carrying a stable ID, typed evidence, and explicit dependency, support, and contradiction edges. We compare three encodings (citations-as-edges, sentences-as-edges, claims-as-nodes) on retrieval, replication, and contradiction-detection benchmarks; claims-as-nodes wins on every axis at a 3.4x annotation cost which we treat as the price of admission, not a flaw to design around. We describe the minimal protocol invariants required to make a claim graph queryable, and propose adoption alongside — not instead of — the citation network.