\section{Conclusion}\label{sec:conclusion}

We set out to learn the scientific writing conventions of a field from a corpus written before conversational LLMs, to build a tool that uses them, and to test the tool against its alternatives under a preregistered protocol. The measurement part succeeded on its own terms: \nIncluded{} \emph{Journal of Finance} articles were verified, parsed and annotated with a schema-validated rhetorical ontology, and the resulting profiles describe, with uncertainty, how empirical and theoretical finance articles organise abstracts, introductions, identification arguments, robustness and tables. The system part delivered a local-first engine that drafts, upgrades, reviews and audits manuscripts from inside Claude Code, never edits a manuscript, and attaches provenance, flags and audits to everything it proposes; it was exercised on the author's own working papers and from a separate project.

The evaluation part gave a clear and somewhat uncomfortable answer. Measured conventions and structurally annotated exemplars, delivered as retrieval, did not improve a capable model's drafts over a strong generic prompt on the dimensions of clarity, logic and positioning; a verification-and-referee stage did, substantially, on rigor and fidelity, and the ablations locate the effect there. The study is small and judged by a model, and we have been careful to say so; but the direction of the result is consistent with what is known about these systems, and it points to where future work should go: human-rated replications on more documents and fields, cost-matched comparisons, referee stages that are themselves audited, and measured profiles for other journals and disciplines so that the question of whether conventions transfer can be answered rather than assumed. The pipeline, prompts, protocols and profiles are released with this paper so that others can do exactly that.