1Large language models now draft, edit and review scholarly manuscripts, but they are usually prompted with generic advice rather than with the measured conventions of a discipline. This paper asks whether abstract, domain-specific writing conventions extracted from a pre-ChatGPT corpus of finance research improve LLM-assisted scholarly writing and review beyond strong prompting and conventional retrieval, while preserving originality and factual fidelity. We assemble \nIncluded{} \emph{Journal of Finance} articles (volumes 72--75, 2017--2020), parse them into sections, equations, tables, figures and references with a layout-aware extractor, annotate every document with a schema-validated rhetorical ontology (abstract and introduction moves, contribution placement, identification framing, robustness and table conventions), and turn the \nTrain{}-document train split into measured writing profiles with bootstrap uncertainty. The profiles and \idxUnits{} provenance-bearing structural units feed JoF-250, a writing, review and audit engine integrated into Claude Code through ten slash commands and seven specialist agents. A preregistered experiment on held-out articles compares five conditions, from a bare prompt to the full engine, with a blind LLM judge and deterministic fidelity audits. Measured profiles and structured rhetorical retrieval do not, by themselves, improve judge-rated clarity, argument logic or contribution positioning over a strong prompt (H1, H2 not supported). The full engine's verification-and-referee stage does: methodological rigor rises by \bmHEvsDMethodologicalRigorDiff{} points and fidelity by \bmHEvsDFidelityDiff{} points on a seven-point scale (Holm-adjusted $p=\bmHEvsDMethodologicalRigorP$ and $p=\bmHEvsDFidelityP$), new numbers absent from the source material fall, and ablations locate the gain in the specialist-referee pass rather than in retrieval or profiles. Outputs share at most \bmOverlapOtherMax{} of their eight-grams with any corpus document other than the task's own source. The study is a pilot (two held-out documents, eight tasks, one out-of-domain manuscript) and relies on an LLM judge; we report its limits, release the pipeline, and argue that in LLM-assisted scholarly writing the binding constraint is verification, not style.2