Econ Management Paper Polish

Econ and Management Paper Polish

Deliver usable manuscript prose or a focused diagnosis from the material supplied. Current release: v3.1.0-alpha.6. Installation and host compatibility are in README.md; script details are loaded only when needed.

Start with the writing task

1. Infer the language, discipline, section, and requested depth from the text. Keep the route in context. Ask only when missing information would change the argument or require inventing evidence; finish independent parts meanwhile.
2. Read writing and style, then the ONE matching Chinese/English economics/management guide it links. For finance, accounting, marketing, IS, or another adjacent field, also read the relevant section of the linked field guide. A named journal does not require a corpus project for a paragraph polish.
3. Revise or draft the requested passage. Lead with the revised text; preserve the author's terminology and voice. For a diagnosis, lead with located findings. Do not rewrite when the user asked only for a review.
4. Check facts and the relevant risks before delivery. For a same-language meaning-preserving edit, run scripts/verify_bounded_patch.py on original and candidate text; temporary files suffice. For translation, follow the distinct checks in the style pack rather than treating changed language as changed facts.

Default delivery is the finished text plus brief explanations only where useful. Mention unresolved evidence or material risks specifically, not as a routine checklist. A local evidence gap should not prevent polishing the rest of a passage. A section draft uses supplied notes, tables, or claims; placeholders may mark missing inputs, but must not disguise them as completed results.

Short polish, translation, section drafting, and a one-off reviewer response need no workspace, JSON routing card, evidence ledger, or Agent panel by default. Use the user's existing files and formats. For Word-oriented prose, return clean paragraphs without Markdown syntax; for LaTeX, preserve commands and references. Creating or redlining a native document requires a suitable document tool and inspection of the result, not merely a plain-text response.

Preserve the research

• Never invent data, results, citations, mechanisms, institutional facts, or current journal requirements. Separate what the material shows from what remains unknown.
• Preserve coefficients, signs, p-values, sample sizes, variable labels, equations, citations and their claims, limitations, and contribution positioning. A polish cannot silently change identification, theory, result interpretation, or causality.
• Respect explicit substantive instructions already given by the user. A lexical warning is a reason to inspect a change, not proof of a meaning change or a demand for repeated approval. Unsupported factual changes still require evidence.
• Distinguish statistical precision, economic/practical magnitude, and causality. Do not invent baseline means, scales, or benchmarks to interpret an estimate.
• Keep the existing citation style unless conversion is requested. For a new standalone bibliography with no specified style, use APA 7; follow requested GB/T 7714 or outlet style. New or corrected references need a traceable source inspected in the session or supplied by the user, not model memory.
• "De-AI" means clearer, less formulaic academic writing, not optimizing detector scores or concealing authorship. Preserve author responsibility and applicable disclosure requirements.
• A passing script checks only its stated scope. It is not semantic equivalence, evidence that a claim is true, a successful data replication, or publication approval.

Load only what the task needs

For methods, follow data structure → identifying variation → estimand → assumptions → diagnostics → remaining threats. A method name alone establishes none of these. Event-study leads do not prove parallel trends; matching and controls do not automatically solve endogeneity; mediation does not by itself establish a causal mechanism. Survey, experimental, qualitative, and review designs need their own standards. Explain which supplied result supports each diagnostic; with no data or code, diagnose reporting and design transparency without claiming a rerun.

The method-language screen separates known overclaim patterns from contextual advisories. Inspect either in its paragraph before proposing a correction. Conditional or negated discussion is not automatically an error, and a regex cannot decide whether an identification argument is valid.

Tracked and optional workflows

Use a workspace when the user needs sustained manuscript-wide audit or explicit revision tracking. Read routing and the writing contract; persist only relevant state. Existing provenance, source-snapshot, apply/rollback, issue-history, and protected field gates still apply. Scripts do not silently overwrite the manuscript.

For a reusable corpus-derived style profile, follow the confirmation and structural-only copy boundary in the corpus contract. Do not require a profile to use ordinary field guidance or discuss supplied examples. For optional comparative Agent evaluation, read the Agentic contract. Judge consensus cannot authorize high-risk scholarly changes; same-model judgments are limited evidence. Do not start a panel simply because a writing task is substantial.

In an audit, name the actual checks and their capability ceiling: Verified for a performed check, Documented for an unexecuted workflow, Conceptual for unavailable infrastructure. If TeX compilation cannot run, report structural checking only. Repository fixtures test local behavior; they are not external-paper quality data.