Research paper 29

Graph-Native Manuscript Development: Claim Bundles, Evidence Anchors, and Verification Reports

Agent-written manuscripts can become polished before they become trustworthy. Graph-native manuscript development offers a different object: sources, claims, sections, citations, draft fragments, review findings, and revisions represented as an evidence graph. This paper studies GraphOrchestra as an early local-first system for that object model. The inspected run attempted to produce a trustworthy literature review of graph-orchestrated manuscript development. Its artifacts include an outline plan, claim bundles, and a verification report. The run is valuable because it did not pass cleanly. The outline reports 146 artifacts, 74 document sections, and 800 code symbols, but no paper nodes and no notebook cells. Claim bundles repeatedly show no citations and narrow source support. The verification report marks the Introduction and Related Work sections as failed, reports warnings elsewhere, and records guardrail violations around evidence support and ambiguity. The contribution is therefore not an automated-manuscript success story. It is a methods and negative-evidence paper: graph-native manuscript systems are promising when they make unsupported claims, citation gaps, numeric mismatches, polarity contradictions, and unresolved ambiguities visible before prose is treated as final.