Citation graphs are not knowledge graphs

Abstract

Citation graphs (paper$\to$paper, untyped, append-only) and knowledge graphs (entity$\to$entity, typed, revisable) are routinely conflated in scholarly-infrastructure discussions, but they make incompatible structural commitments. We enumerate six structural differences and document the concrete failure modes that follow when one is treated as the other — for example, running KG entity-merge logic over a citation network corrupts the citation chain in ways that retraction handling cannot recover. We then propose a minimal typed-edge extension to the citation graph (depends_on / supports / contradicts / extends) that recovers roughly 89% of the queries a KG baseline could answer while remaining compatible with BibTeX-shaped tooling. The extension is implemented in the rrxiv reference server and round-trips through cir.schema.json. This paper is the formal counterpart of the claim-graph proposal in rrxiv:2605.00002: claim-graph operates at the claim level; here we operate at the paper-and-claim level with typed edges between them.