Research Paper Writing with Rigor
Write a paper that lets a knowledgeable reader see the result, understand why it follows, identify what is new, and obtain the evidence needed to check it. Use direct language for established facts and precise uncertainty for open questions. A short note earns its concision through a focused question, removal of repetition, and clear organization. Every core conclusion it retains needs a complete argument.
Primary principle: make the completed contribution visible
The paper's main job is to explain what the authors accomplished, how they accomplished it, and why it matters. Organize the abstract, introduction, result summaries, and conclusion around those completed contributions. Give supported results their full strength: "we prove", "we construct", "we establish", and "we formalize" are appropriate when the evidence warrants them.
Excessive modesty can misrepresent the work as surely as exaggeration. Remove "merely", "only a modest step", "we attempt to", or "may suggest" when they weaken an established result without a scientific reason. Preserve qualifiers that express actual uncertainty, statistical evidence, or mathematical scope. Neither routine attribution nor an unresolved extension requires an apology for the present result.
Evaluate a proposed caveat by the reader's needs: would omitting it materially change the claim's meaning, evidence level, applicability, or attribution? If yes, state it precisely at the relevant point. If no, remove it or move useful operational detail to the supporting documentation. Do not invent hypothetical claims a reader might make merely to deny them. Avoid opening or closing a contribution paragraph with a catalogue of things the paper does not do.
Use evidence checks to earn confident writing. A review finding should lead to the strongest accurate statement supported by the work; default hedging is not the goal. The final reader should remember the result and its mechanism.
For a new draft or substantial revision, apply these checks in order:
- State the question and main result in the first paragraph of the introduction; do the same within the abstract if there is one. Give the actual scope, using a faithful plain-language version when formal notation must come later.
- Keep material that answers the question, proves the result, explains its mechanism, or places it in the literature. Cut unrelated claims and repeated narrative across the abstract, introduction, body, and conclusion.
- Give every retained mathematical conclusion a complete proof or a precise, checked reference with its hypotheses matched here. Remove an unsupported conclusion from the results; hedging cannot supply its missing argument.
- Replace routine defensive phrases with the supported statement, attribution, or one relevant open question. State mixed verification scope once in a status sentence or table; preserve the hypotheses and finite premises in the statements that need them.
- Cover the literature systematically. For a short paper, plan 12–20 or more relevant, verified references; follow the coverage and verification checks in literature-and-citations.md.
- Finish with a read-aloud pass for rhythm, then recheck mathematical meaning after any sentence changed for fluency.
Read focus-and-structure.md for the scope test, repetition audit, 2–4 page note and short-paper templates, and final acceptance checklist. Apply these to the requested passage for a local edit; do not turn it into an unsolicited full-paper audit.
Choose the requested work
- For a sentence or paragraph, give the concrete revision and only material explanations. Do not create a manuscript-wide audit or release workflow.
- For drafting or substantial revision, read the current source, relevant results, and coauthor comments before editing. Work from the actual latest version, not a remembered draft or stale line numbers.
- For review only, lead with findings and locations; do not silently rewrite files. Distinguish a demonstrated error from an unresolved concern.
- For a submission package, read release-and-reproduction.md. Editing prose alone does not authorize repository publication, journal or arXiv submission, messages to others, or shared CI changes.
Adapt the structure to the discipline, venue, and requested length. A short mathematical note need not have IMRaD sections, ablations, a teaser figure, or an experimental-performance narrative. Follow explicit user instructions and applicable venue requirements over these defaults.
Learn the discipline's written language
Treat academic style as a set of field-specific explanatory choices: which objects are named, what can be assumed, where an argument needs a calculation, and how results are related to earlier work. Formal-sounding vocabulary alone does not supply these choices. For substantial drafting or complaints about generic or machine-like prose, read discipline-and-voice.md.
Before writing technical content, search recent relevant papers in the field and, when applicable, the correspondent's own papers. Read the relevant passages and adopt their standard terms and notation; when writing to an author about their paper, use that paper's notation. For small edits, reuse already checked sources when sufficient. Learn the field's organization and level of explanation while preserving the argument and authors' voice. Do not imitate an author's identity or copy distinctive phrases.
Use established field terms; do not coin a name when a standard term exists. If a new term is unavoidable, define it in one sentence at first use and relate it to the standard term or object. Prefer direct theorem statements; add an optional theorem label only if readers will use it. Avoid naming every observation or importing internal research labels. A proof paragraph should expose the inference that makes the conclusion follow; editing words cannot replace a missing inference.
Judge the revision by its usefulness to a specialist, mathematical fidelity and fit with the authors' prose. Do not optimize for detector scores, invent human drafting history, introduce errors or slang to seem human, or promise that the text's origin will be undetectable.
Establish the argument before polishing
Identify the question, main result, key obstacle, proof or method mechanism, closest prior result, and remaining uncertainty. For long work, use a short private outline or claim-source map if it helps; reuse existing records. Do not insert drafting labels or audit tables into the manuscript by default.
For substantial mathematical-note drafting, coauthor revisions, or feedback that proofs are too compressed, read proof-completeness.md. Check the actual inference chain before shortening it: changes of representation, normalization, dimension, summation order, and boundary cases need their mathematical reasons. Keep the main argument readable and place longer necessary derivations in precisely referenced appendices. Length is a constraint on scope and exposition, not evidence that a proof is complete.
For substantial revision, check which auxiliary results and appendices the main argument actually uses. Keep an independent proof or example when it provides a distinct result, method, obstruction, or explanatory benefit worth its length. A consequence already supplied by the main theorem may need only its deduction. Before removing material, preserve every dependency and update the abstract, introduction, cross-references, and bibliography.
For each central claim, know its scope and supporting evidence. Preserve quantifiers, initial indices, exceptional cases, hypotheses, and constants. Separate new mathematics from a new proof, a formalization, an implementation, and an empirical observation. Do not equate an agent's agreement or a finite prefix match with an all-input proof. Read argument-and-evidence.md for proofs, computer-assisted results, literature placement and application, or mixed verification. Its literature review includes a dedicated independent referee pass and a hard documented search gate before calling a manuscript or short note ready.
Call something a "counterexample" only when it falsifies a precisely quoted statement: quote that statement, check its hypotheses, and identify the failed conclusion. Otherwise call it an example or illustration and say what it shows.
If support is missing, complete the argument, prove a narrower statement, or remove the claim and its dependent assertions. Keep a conjecture or open question only when it serves the paper, clearly separated from established results. Do not retain an unsupported conclusion by adding "may", "appears", or a disclaimer. Do not invent proof steps, references, numbers, author actions, model identities, successful checks, or public artifact links. A blocked claim need not stop independent writing work on supported results.
Write as a research paper
- Lead with the mathematical or scientific content: what is established and the mechanism that establishes it. State the contribution specifically and at its full supported strength; keep it as the paragraph's main message.
- In the introduction, say why the question matters before the proof: the difficulty it resolves and the consequences the result makes available. Give each paragraph a clear job. Check the sequence of paragraph messages before polishing sentences. Connect definitions, arguments, evidence, and consequences in the order a reader needs them.
- Use precise verbs and stable terminology. Prefer an identifiable object, such as "the correction automaton E", to "the supplied correction". Avoid internal project language and chronology unless they matter to the result. Repeating a defined term is preferable to a misleading synonym. Check what each symbol denotes, its domain, and its normalization. Make the map and measure clear when moving between representations or writing integrals. Define a local use of a term that also names a different standard quantity in the field.
- Place the contribution in the wider literature and credit the specific method or theorem where it is used. Replace repeated novelty denials with positive attribution and a clear statement of this paper's extension. A citation must support the exact attached claim.
- Separate indispensable limitations from defensive repetition. Keep a qualification beside a claim if omitting it would change its interpretation; consolidate verification status in one sentence or table. Use a limitations section when it has substantive content, not as a ritual closing paragraph.
- State open problems as open problems. Distinguish an unproved identity, minimality, attainment, asymptotic behavior, and a dynamical interpretation. Do not replace "not established" with "false" or "unnecessary".
- Put operational details where readers need them: essential proof evidence in the paper or a precise accessible reference, reproduction commands and detailed environment records in the supporting package. The main text should not become an account of internal orchestration or mail handling.
Read editing-patterns.md when tightening prose or applying coauthor comments. Its examples are decision patterns, not mandatory sentences to paste into every manuscript. Do not mechanically ban passive voice, hedging, repeated terms, or punctuation that serves the argument.
Apply comments with judgment
Classify material comments as correctness, attribution, scope, reproducibility, or presentation. Check technical replacements against their surrounding proof; do not infer validity from the reviewer's authority alone. A shorter sentence must retain the witness bounds or hypotheses that make it true. Verify removal of an algebraic assumption with exact reasoning where feasible.
A precise citation can replace a proof of a standard tool. Read the cited statement and match its domain, topology, boundedness or integrability, and normalization requirements to this application. Keep the instantiation and the conclusion used here explicit. When shortening local reasoning, retain the bound or hypothesis that licenses the next inference. When a known existence theorem supplies an exact optimizer, a proof can often use that optimizer instead of an approximate one with a split error budget; verify the theorem's hypotheses first.
Implement supported changes when revision is requested. Explain disagreements with evidence and propose a precise alternative. When a coauthor's supervisor comments, address every point concretely in a change log: comment, manuscript location, change or reasoned response, and verification or unresolved issue. For other reviews, keep a compact disposition record when the number or importance of comments warrants one. Keep this log outside the manuscript; do not require a separate approval checkpoint for every reversible edit.
When following several revisions, distinguish the received draft, the corrected draft, and any later published version. Track important passages by their mathematical content as numbering and notation change. Recheck that previously accepted corrections remain effective; an equivalent reformulation may preserve a correction without preserving its wording.
For feedback about formulaic prose or excessive theorem labels, read the comment and affected passage together. Identify the concrete defect, revise the passage, and extract a reusable lesson only if it applies beyond that example. Recheck the rendered statement and its proof dependencies; changing a heading alone does not address missing explanation. Use the focused review in discipline-and-voice.md.
Short result notes for a correspondent
A short note answers one reader's question first: the author whose open problem, conjecture or request it addresses. Read their exact statement (paper version, section, equation) and answer that; move anything else to one closing remark or cut it.
Follow the requested or venue length rather than imposing a fixed page range. Write the complete argument for the retained claims before compressing its presentation. Under a tight page limit, remove secondary claims or use allowed appendices before deleting a necessary inference. Standard tools can be cited with their applicable hypotheses; a central new step needs a written proof.
- Before writing, list each claim with its status: proved here, known (exact source), known in a special case (state precisely what is new), computed (finite range and method), or open. Only proved claims become theorems.
- Title states the mathematics, never "Response to X". Use the byline and affiliations the task supplies. Abstract gives the main statement exactly and its scope in three to six sentences.
- Introduction: the question as the source states it, the answer, its relation to known results. Then setting in the source's notation and results with complete proofs. Include examples, certificates, or open problems when they add information; finish with references. Use the templates in focus-and-structure.md at the requested length.
- Keep an attribution ledger for a correspondent-facing note: identify which results are proved by the present authors, which are independent checks of another author's result, and which are direct consequences of cited theorems. Describe the latter two as verification or consequences, never as advances.
- Computation supports but never replaces a proof of an all-parameter statement; give finite ranges explicitly. Never cite your own unpublished drafts; fold what is needed into the note. Credit the correspondent's question, examples, corrections and counterexamples plainly in ordinary prose, for example “in correspondence, 19 October 2026”; do not create an email bibliography entry. Never cite or mention "private communication", including hyphenated variants and self-certifying sentences such as "no private-communication references".
- Keep process and evidence logistics in a repository README, not the note. No local paths, no tool or pipeline narrative, no long hash strings.
- Compile, inspect every page, recheck every number against its computation, and record what is proved, computed and open, plus any citation that could not be verified.
For delivery of a research note by email, package one merged PDF named note_YYYYMMDD.pdf and one dated source ZIP (for example verification_YYYYMMDD.zip), with total attachments <= 5 MB (5,000,000 bytes). Fold every meaningful result into that manuscript or note. The coauthor package has one deliverable: the manuscript PDF and at most one source archive. Put verification material in the public repository at a pinned link instead of adding separate result notes or extra archives. The ZIP contains text/source and concise readable outputs only; exclude binary files (.bin, .dat, .npz, .gz, .pkl), third-party copyrighted PDFs, and internal RESULTS-r*, REVIEW-r* and NOTES files, including nested copies. The merged note PDF is the document-format exception. Omit regenerable generator output even when it is text: give the exact command and output SHA256 in the package README. Large data belongs in the authors' designated public data repository; use a verified commit-pinned link and do not invent or publish artifacts without authority. Check final PDF text, source and archive members for long digit runs, including space/hyphen-separated digits, that pass Luhn and trigger institutional DLP. Use exact symbolic forms, concise scientific notation or pinned public data links to keep those runs out of delivered material.
When asked to review such a note adversarially, recompute independently, read each proof line by line, verify every citation against the primary source, look for scope overreach and missing hypotheses, deliver a corrected version and a verdict: correct, fixable gaps (listed), or wrong at a named step.
Verification, authorship, and AI use
Report the actual verification scope: which statements were proved, which were kernel checked, which computations ran, and which results are merely reported by another author. A successful build is evidence about a specific source version; it is not validation of all prose or a fresh independent review.
When the authors have a prescribed AI-use sentence, use it verbatim; do not substitute another sentence or add a process narrative. Do not attribute that use to other collaborators. Keep detailed provenance and actual verification scope in internal research records. Do not claim every proof was kernel checked when only a subset was formalized, or imply every coauthor reran the tools or approved a pending draft. Answer direct process questions truthfully. If a venue requires additional disclosures, flag the specific requirement separately rather than silently changing this sentence. Do not insert these authors into unrelated manuscripts; establish their actual authors and applicable disclosure requirements from the task. Keep data/code availability in a separate statement; see release-and-reproduction.md.
Distinguish permission to use AI in research from permission to use AI-written submission prose. If the selected venue requires the actual wording to be human-authored, supply mathematical derivations, source checks, an outline and comments on the humans' text in the permitted roles. Keep model-written examples as working material; they do not become human-authored by paraphrasing, translation or a final approval pass. Apply this boundary to that venue rather than prohibiting manuscript drafting for every other venue.
Finish proportionately
Before sending or submitting a draft, complete the independent terminology and readability referee pass in discipline-and-voice.md.
Check that the abstract, introduction, theorem statements, tables, conclusion, and verification description agree on the central claims. Read the abstract and conclusion once just for emphasis: do they make the completed result and its significance clear, or does a list of exclusions obscure them? Correct unnecessary hedging as well as overstatement. Run the cross-section repetition audit and final read-aloud pass in focus-and-structure.md. For changed LaTeX, compile the affected deliverable when available and inspect relevant rendered pages; report unavailable validation honestly. Do not rerun expensive formal jobs solely for prose changes unless changed claims or project rules require it. Respect the configured validation machine and existing repository gates.
When assessing concision, compare like scopes: report whether the reduction comes from removed auxiliary material or from a shorter treatment of the retained argument. State the counting convention when lengths matter, and check any reported page count against the actual rendered artifact.
Deliver the requested prose or artifact first, then a short account of material changes, completed checks, and remaining decisions. Keep internal checklists out of the publication text. Do not promise acceptance, universal novelty, or that all possible objections have been eliminated.
For provenance and deliberate departures from upstream writing skills, read sources.md. This skill contains paraphrased lessons from a private mathematical coauthor review, not the correspondence itself.
CHANGELOG
- 2026-10-29 — 2.0: make focus, cross-section deduplication, complete support for retained conclusions, and removal of defensive boilerplate explicit acceptance checks. Add note/paper templates, systematic literature coverage with a 12–20-reference short-paper target and index verification, and a final read-aloud pass. Synthesize nine public writing skills and guides with pinned source attribution; retain the proof, authorship, and reproduction guidance.
- 2026-10-29 — 1.10: make proof completeness operational for compact notes: explicit transformation and coefficient checks, domain and dimension tracking, justified boundary limits, equality arguments, and precise appendix links. Preserve coauthor scope and remove the fixed short-note page range.
- 2026-10-29 — 1.9: add a hard, documented literature gate covering recent searches and overlap checks; require minimal manuscript packages with a pinned public verification link; and separate our results from independent checks and cited consequences in correspondent-facing notes.
- 2026-10-19 — 1.8: use the authors' prescribed AI sentence verbatim, prohibit correspondence citations and self-certifying mentions, and require small dated note/source packages with reproducible public data links, no internal process files or binary data, and DLP digit checks.