Code of conduct

Last updated 2026-05-19. Plain English.

This document sets out what OpenXiv expects from people who use the platform. It is short and concrete on purpose. We do not promise utopia; we promise that violations are handled openly, proportionally, and with the work as the unit of judgement, not the person.

For authors

Integrity. The work is yours, your co-authors', or correctly attributed prior work. Fabricated data, doctored figures, hallucinated citations, or invented references are grounds for refusal.
Verification. Every citation, formula, and numerical claim should have been checked by a human author before submission. The full AI-disclosure framework lives in the submission policy; unverified or hallucinated output without that disclosure is grounds for refusal.
Conflicts of interest. Disclose funding, commercial interests, or institutional positions that could bias the work. The disclosure goes in the manuscript; we do not run separate forms.
Versioning, not silent edits. If you find an error after publishing, submit a new version with a changelog. Don't request that the original be quietly modified.
Engagement with critique. Endorsers and readers may raise issues. Engage honestly. Disagreement is fine; ignoring substantive critique while leaving the work unchanged is bad faith.

AI use: explicit rules

AI tools (large language models, code assistants, image generators) are welcome on OpenXiv under clear, concrete rules that match the consensus of major journals (Nature, Science, JAMA) and publication-ethics bodies (COPE). The goal is not to discourage AI use; the goal is to make AI use legible, verifiable, and accountable.

AI is never an author. Authors are humans (or organizations through human representatives). AI tools used in producing a paper are disclosed in the AI-disclosure record, not the author list. Listing an AI agent as a co-author is grounds for refusal.
AI involvement is disclosed at one of four levels. Every submission tags its AI use as none, assistant, co-author, or primary; names the models used (e.g. "Claude Sonnet 4.5", "GPT-4o", "DeepSeek V3"); and describes briefly how each was used. Misrepresented levels are grounds for refusal.
Every citation is human-verified. AI tools regularly fabricate references that look real. Every citation in the manuscript must have been checked by a human author against the actual source. Hallucinated citations are grounds for refusal.
Every numerical claim is human-verified. AI tools regularly fabricate computations, statistics, and numerical results. Every quantitative claim must have been reproduced or independently verified by a human author. Fabricated numbers are grounds for refusal.
No untouched LLM boilerplate. Phrases like "as an AI language model", "I apologize for the confusion", or "I cannot do that, but here's" have no place in a research manuscript. Their presence indicates the output was never edited by a human, and the submission is refused.
The author owns the work and the errors. AI use does not transfer responsibility. The human authors attest to the correctness of every claim in the submission, regardless of whether AI helped produce it. Disagreements about correctness are addressed through endorsements, versioning, and refusal packets — not by blaming the model.

These rules apply to text, code, figures, tables, and any other content produced with AI assistance. They are content-neutral: the standard is "verify before you submit", not "avoid certain topics".

For endorsers

Typed endorsements are statements of fact. "Reproduced the result", "checked references", "verified derivation": each one means you did the work and stand by it. Casual or transactional endorsements (you-endorse-mine-I-endorse-yours) corrode the signal that makes typed endorsements worth anything.
Scope. Endorse what you actually verified. If you reproduced a numerical result, that is one endorsement; if you did not check the derivation, do not issue "verified derivation". Specificity is the value.
Conflicts. Do not endorse work where you have a financial, institutional, or close personal stake. Co-authors, supervisors, and direct collaborators are excluded by default.

For everyone

Argue the work, not the person. Critique of a paper, method, or claim is welcome. Personal attacks, harassment, doxxing, or threats are not.
No protected-class harassment. OpenXiv is for research. Behaviour targeting people on the basis of race, ethnicity, religion, gender, sexual orientation, disability, or nationality gets a hard refusal.
Privacy. Do not share other people's private correspondence, personal data, or identifying information without consent.

For moderators

Refuse the work, not the author. Every refusal is for a specific submission, not a person. Authors may resubmit a revised version. We do not issue year-long bans for honest mistakes.
Public refusal packets. When a submission is refused, the refusal packet is public: it names the failure mode, points at the evidence, and recommends a revision path. See the submission policy for the standard format.
Proportionality. A single hallucinated citation is fixable. A pattern of fabricated data across the manuscript is a hard refusal. We err toward "return for revision" when intent is unclear.
No silent moderation. Every action affecting visible content is logged on the transparency report.

Sanctions

OpenXiv's ladder of responses, from gentlest to firmest:

Return for revision. The submission has issues; the author can fix and resubmit.
Public refusal packet. The specific submission is refused. The author may submit different work later without prejudice.
Endorsement removal. An endorsement found to be fabricated or transactional is removed; the endorser is notified, and the removal is logged publicly.
Account suspension. Reserved for hard cases: harassment, repeated bad-faith submissions after refusal, or account compromise. Suspensions are documented on the transparency report. Appeals are public and timed.

We do not have permanent author bans for honest mistakes. We do have permanent removals for confirmed plagiarism, doxxing, or systematic fraud.

Reporting

To report a code-of-conduct violation, email davidich.alfyorov@gmail.com with the specific item, the concrete issue, and any evidence. We acknowledge within 7 days. Outcomes are logged on the transparency report, redacted only for reporter privacy where requested.

Changes to this document

Material changes (new sanction categories, scope expansion) are announced at least 30 days before they take effect. Typographic or clarifying edits do not require notice; the current text of every policy is always at the URL above.