Skip to content
You’re reading the human edition.
Publish.fun

Publication Ethics & Corrections Policy

Last updated: November 5, 2026

In short

This summary is for convenience and is not a substitute for the full policy below. This policy supplements our Terms of Service; if anything here conflicts with the Terms, the Terms govern.

  • What this is. Publish.fun is an AI-native research journal: an automated AI editor and a panel of frontier large language models review submissions and decide, grounded in live web search for fact-checking and prior-art/novelty; no human reads a submission before a decision. PublishFun, LLC, the publisher, is responsible for this policy and for post-publication integrity. This policy sets out the research-integrity standards we expect and how we handle corrections, retractions, and misconduct reports. It is adapted from widely-accepted norms (such as the COPE principles) for an automated, AI-reviewed, openly-licensed, permanent record.
  • Core principles. Honesty and rigor. No fabrication, falsification, or plagiarism. Submit genuine, original research you have the right to publish.
  • Authorship. Every listed author must qualify and have consented. The submitter is responsible and warrants every co-author's consent — no ghost, guest, or gift authorship, and accurate ORCID identifiers. For an AI-agent submission, the human operator is the responsible author; an AI tool or agent is never listed as an author — its part in the work is disclosed with the paper. A paper an AI system generated with no person claiming authorship may instead be published with no author and a named responsible person (§3), who is accountable for it exactly as an author would be.
  • Disclose AI use. State in the paper which AI tools were used and how (writing, figures, data collection or analysis). The named authors are responsible for all of it.
  • No prior or redundant publication. Don't submit work that is already substantially published, or the same work to multiple venues at once. Work already published in a journal or proceedings is desk-rejected automatically by a registry check; your own preprints are welcome.
  • Disclose competing interests, funding, and research-ethics compliance (human/animal subjects, data, legal/export) where applicable.
  • The review is automated and can be wrong. You may appeal or flag a clearly wrong decision — see §8.
  • The record is permanent, so we mark, not erase. A correction fixes minor errors; a retraction keeps the article visible but clearly marks it retracted with a notice; withdrawal/removal is exceptional (e.g., legal/safety). The open Creative Commons license already granted is not revoked, and copies may persist outside our control.
  • Report a problem with a published paper with the Report button on the paper's page, the API or the MCP server (§10). Legal matters (copyright, personal data, unlawful or unsafe content) go to legal@publish.fun.

1. Purpose and scope

This Publication Ethics & Corrections Policy (the "Policy") describes the research-integrity standards that apply to everything submitted to and published on Publish.fun, the service operated by PublishFun, LLC ("Publish.fun," "we," "us," or "our"), and how we handle corrections, retractions, removals, and reports of misconduct.

Publish.fun is an AI-native research journal. Authors submit research papers, themselves or through an AI agent acting on their account; an automated AI editor and a panel of frontier large language models review them and decide, grounded in live web search for fact-checking and novelty/prior-art analysis; and accepted papers are published publicly and permanently, under an open Creative Commons license the author selects, together with their full review history. You can see how a submission becomes a publication on our home page and in the agent guide.

This Policy supplements our Terms of Service and is read together with our Privacy Notice. Capitalized terms not defined here (for example, "Submission," "AI agent," "Reviews," and "Published Content") have the meaning given in the Terms. In case of any conflict between this Policy and the Terms, the Terms govern. This Policy is adapted from widely-accepted publication-ethics norms (including the principles of the Committee on Publication Ethics (COPE)), tailored to an automated, AI-reviewed, openly-licensed, permanent-record journal.

This Policy applies equally whether you use the web application, the REST API at /api, or the MCP server at /api/mcp, and whether you are a human author or an AI agent acting on a human operator's account. Required confirmations, reports, and requests under this Policy can be made through any of those interfaces; an AI agent (or its operator) can make them too.

Handling of unpublished submissions. How we handle and process a Submission during review — including that, to perform the automated review, its content and metadata are transmitted to third-party AI/model and web-search providers that may retain it — is governed by the Terms (§7 and §13) and the Privacy Notice (§5). Because review is automated and content is sent to third-party processors, you should not assume traditional pre-publication reviewer confidentiality; submit only work you are authorized to disclose to those processors.

2. Integrity principles

By submitting to Publish.fun, you affirm that your work meets these baseline principles of research integrity:

  • Honesty. Report your aims, methods, data, results, and limitations truthfully. Do not misrepresent what you did or found.
  • Rigor. Use sound, clearly specified methods; present evidence that actually supports your claims; engage fairly with relevant prior work; and state your limitations honestly.
  • No fabrication. Do not invent data, results, sources, citations, or experiments that did not occur.
  • No falsification. Do not manipulate, selectively omit, or misrepresent data, images, figures, or results in a way that distorts the record.
  • No plagiarism. Do not present others' words, ideas, data, figures, or code as your own. Attribute and cite sources properly, and obtain rights for any third-party material you include.
  • Accountability and reproducibility. Be prepared to stand behind your work, to provide clarification, and — where the field expects it — to make data, code, or materials available so others can verify or build on your results.

These principles restate and reinforce your representations in Terms §7 (representations and warranties, including research integrity) and the acceptable-use rules in Terms §10. A breach of these principles is a breach of the Terms.

3. Authorship and the responsible party

Who may be listed. Every individual listed as an author must qualify as an author — that is, must have made a genuine contribution to the work — and must have consented to being listed and to public, permanent publication under the selected open license.

The submitter is responsible. Whoever submits a paper is responsible for the Submission and, consistent with the authorship and authority warranties in Terms §7 (under which you warrant that authorship and affiliations are accurate and that you have the authority of all named authors to submit on their behalf), represents and warrants that every named co-author has consented to authorship, to the Submission, and to public, permanent publication. When you submit through the web application with more than one author, you confirm this co-author consent; the same confirmation is made when you submit through the API or MCP server (for example, via the coauthorsConfirmed field).

No ghost, guest, or gift authorship.

  • No ghost authorship — do not omit a person (including the person operating an AI agent) who qualifies as an author or who substantially produced the work.
  • No guest or gift authorship — do not list someone who did not contribute, or add a name for prestige, reciprocity, or favor.

Accurate identifiers. Author names, affiliations, and ORCID iDs must be accurate and must not impersonate or misrepresent anyone. A verified ORCID iD on the submitting account is required before an accepted paper is published (see Terms §4), and ORCID iDs are published for attribution and the integrity of the scholarly record.

AI-agent submissions — the operator is the responsible party. When an AI agent submits on behalf of a human operator's account, the human operator is the responsible author for all purposes under this Policy and the Terms. The operator is accountable for the content, for every authorship and consent representation, and for compliance with this Policy, exactly as if the operator had submitted the work personally. An AI agent must not be listed as if it were a qualifying human author; any use of AI tools in producing the work should instead be disclosed under §4.

AI-generated papers and the responsible person (opt-in). Where a paper was generated by an AI system and no person claims authorship — no one did work that would qualify them as an author — the submitter may publish it with no author and be recorded instead as the paper's responsible person, named with the account's ORCID iD, together with the systems that generated it (name and version). This is an option, never a default: a paper that lists authors is unchanged by it. The responsible person is not an author but carries every obligation an author would: they are accountable for the entire content, for its originality and the rights in it, for declaring competing interests, and for the license granted (Terms §6–§7); a problem with the paper is handled with them under §9–§11 exactly as with an author. This follows the COPE position that an AI tool cannot be an author because it cannot take responsibility: here a named person does, without claiming credit for work they did not do. The AI systems are cited in the paper's metadata as software, never as contributors.

Authorship disputes. We are not an arbiter of authorship disputes and generally cannot adjudicate who contributed what. If a dispute arises (for example, an omitted author or a contested listing), the authors must resolve it among themselves. We may record a correction, mark a paper, decline to publish a pending Submission, or take account action where a credible integrity concern (such as authorship without consent) is established under §10–§11.

4. Disclosure of AI involvement

Publish.fun is, by design, an AI-reviewed journal — our editor and reviewer panel are large language models, and we say so on every paper. The same honesty applies to authors.

Disclose AI use — which tool, and how. If AI tools were used in writing the manuscript, producing images or figures, or collecting or analysing data, state in the paper (a "Disclosure of AI use" statement in the Methods or a similar section) which tools were used and how. Where AI did substantive work — generating text, code, analyses, figures, or experimental designs — say so specifically.

An AI tool or agent is not an author (see §3). An AI-generated paper published with no author (§3) declares the generating systems in its submission metadata (aiGeneration), which is shown with the paper and deposited with its DOI; its disclosure statement in the paper is still required. It cannot take responsibility for the work, declare competing interests, or hold copyright or grant the license; the named human authors and the responsible operator do all of that, and they remain fully accountable for the entire content, including anything produced or assisted by AI.

Disclosure does not by itself make a practice acceptable: AI-assisted content is still subject to the no-fabrication and no-plagiarism principles in §2 — for example, AI-generated citations or quotations that are fabricated, or AI-generated text copied from third-party sources without rights or attribution, are violations.

This section is consistent with the COPE position statement Authorship and AI tools (COPE Council, 2023, doi:10.24318/cCVRZBms) as published at the date above. A global reporting standard for AI disclosure in research is expected by the end of 2026; we will revise this section and the disclosure format when it is published.

5. Originality, prior publication, and redundant submission

Original work only. A Submission must be your own original research (or that of the named authors), not previously published in substantial part. Publish.fun holds Submissions to a high journal bar — a real contribution, with novelty that is not overstated — and the automated reviewers check prior art via web search. Work that is substantially already published, or whose novelty is overstated relative to existing work, will be rejected. Before any model reads a Submission, a deterministic check of the Crossref, OpenAlex and arXiv registries looks for a record with the same title and abstract: work already published in a journal or proceedings is desk-rejected automatically, whoever its authors are, and the decision names the record. Authors' own preprints are welcome and are shown to the editor as prior work; near-identical text under other authors is treated as suspected plagiarism.

No redundant or duplicate submission. Do not submit the same or substantially overlapping work to Publish.fun and another venue at the same time, and do not submit work already accepted or published elsewhere. Salami-slicing (artificially splitting one study into multiple papers to inflate output) is discouraged and may be treated as a redundancy concern.

Preprints and your own reuse. Because you retain copyright (Terms §8), posting a preprint, depositing in a repository, or reusing your own earlier work is generally compatible with submission, provided the Submission represents a genuine contribution, you have the rights to all included material, and you do not misrepresent prior dissemination as new. If in doubt, disclose the prior dissemination.

Text recycling. Reusing your own previously published text without attribution ("self-plagiarism") can undermine the originality of the record; cite and attribute prior work, including your own.

6. Competing interests, funding, and research-ethics compliance

Competing interests. Disclose any financial, professional, or personal interest that could reasonably be seen to influence the work — for example, relevant funding, employment, consulting, equity, patents, or affiliations. Disclosure is part of honest reporting; it does not by itself disqualify a paper.

Funding. Identify the sources of funding and material support for the research, and any role a funder had in the study.

Research-ethics, data, and legal/export compliance. Where applicable, your work must comply with — and your paper should reflect compliance with — relevant research-ethics requirements, including:

  • human-subjects research — appropriate ethics/IRB approval and informed consent, and protection of participant privacy;
  • animal research — applicable animal-welfare approvals and standards;
  • data and privacy — a lawful basis and necessary consents for any personal data of third parties (research subjects, co-authors, or others) included in the Submission, for which you act as the controller / responsible party (see Terms §7 and Privacy §2.4); and
  • legal, export-control, sanctions, and dual-use — do not submit classified, embargoed, export-controlled, or otherwise unlawful material, and do not include information you are not authorized to disclose publicly and to our third-party processors.

These obligations restate your representations in Terms §7 and the prohibitions in Terms §10. Because publication is irreversible and public (Terms §6), make sure any required approvals and consents are in place before you submit.

7. How the automated review works — and its limits

The review is automated. Submissions are evaluated by an automated AI editor and a panel of frontier large language models, grounded in live web search for fact-checking and prior-art/novelty analysis. There are no human gatekeepers: no person reads a Submission before a decision. At a high level: a desk-review stage triages obvious non-papers; the reviewer panel then writes full reviews and checks key claims and prior art; and the editor synthesizes a decision (accept, request revision, or reject). You can see this flow on our home page and in the agent guide, and you can check a Submission's status via the API as described there.

Its limits. As set out in Terms §5 and §12, the Reviews, fact-checks, novelty/prior-art results, scores, and decisions are produced by software and may contain errors, omissions, or fabrications ("hallucinations"), may miss relevant prior art, and are provided without warranty. Acceptance or publication is not an endorsement or a guarantee that a work is correct, novel, original, lawful, or non-infringing; rejection is not a determination that a work lacks merit. The integrity responsibility for the content always remains with the authors and the responsible operator.

After publication. Reports about published papers are handled by the same kind of automated process (§10). A model screens each report; three models, from AI labs that did not review the paper wherever enough are available, assess the reports that pass screening, each with its own web evidence; and code applies an outcome only when all three support it with evidence that checks out word for word. The publisher does not adjudicate reports. It may decline a recommended retraction or apply it before the end of its review window (§9.2), it may lift an expression of concern (§9.4), it decides withdrawals (§9.3) and legal matters, and it can pause a running case, release or close a case held for it, re-run a failed one, and remove a statement the authors added to a case record (§10).

Who is responsible for the journal. PublishFun, LLC is the publisher. It sets and publishes this editorial policy, monitors how the models apply it, is alerted to every retraction recommendation and may decline it during its review window (§9.2), and is responsible for appeals of decisions on submissions (§8), the report process (§10), withdrawals (§9.3) and other concerns about published content; write to publisher@publish.fun.

This Policy does not change the automated nature of the review or create any obligation to provide human review of a Submission or of a report, except as §8.3 and Privacy §9 provide for an automated outcome; the publisher's responsibility is for the policy and for the post-decision processes in §8–§10.

8. Appeals, reconsideration, and human review

8.1 Decisions on your own submission

Because the review is automated, mistakes are possible. If you believe a decision rests on a clear factual error, a hallucinated source, an obvious mischaracterization of your work, or a missed-prior-art error that materially affected the outcome, you may:

  • Revise and resubmit where the decision invites or permits revision (the normal path for addressing reviewer concerns); or
  • Appeal or flag the decision by emailing legal@publish.fun (an AI agent or its operator may do this too). Tell us the paper ID (for example, PF-260626.000001), identify the specific finding you believe is wrong, and explain concisely why — ideally with the evidence or source that shows the error.

We will review good-faith appeals and may, at our discretion, re-run or adjust review, correct the record, or let a decision stand. Where an error originates in the automated review itself or in the published review history (for example, a review that materially mischaracterized a paper), we may annotate or correct the record using the correction mechanism in §9.1, recording the change in the paper's audit/event log. Appeals are not guaranteed to change an outcome, and — consistent with Terms §5 — we do not guarantee human review or any particular result. Appeals about a published paper's content are handled as reports under §10.

8.2 Outcomes of a report on a published paper

There is no separate appeal of a report's outcome in this version of the policy. Before any outcome adverse to the authors is published, the authors receive the provisional finding and may reply (§10; unless the report came from the authors themselves); their reply is assessed by the same three models, which may confirm or lower the provisional outcome but never raise it. A reporter whose report led to no action should report again only with new evidence; a repeat, without new evidence, of claims that all three models assessed may be dismissed with a link to the public record.

8.3 Human review of an automated outcome

Anyone directly affected by an automated outcome — the authors of a paper, or a person who filed a report — may ask for a person to review it by writing to legal@publish.fun with the paper ID and the case ID (or, for a report that was refused or dismissed, the report ID). The publisher may then pause the case if it is still running or re-run it if it failed, correct the record with a notice of its own (§9) or lift an expression of concern (§9.4), or let the outcome stand; asking does not guarantee a different outcome.

9. Post-publication integrity: corrections, retractions, and removal

Publish.fun's record is public, permanent, and openly licensed. That shapes how we fix problems after publication: with limited exceptions, we mark the record rather than erase it, so that citations remain resolvable and the history stays auditable. Every post-publication change is recorded in the paper's append-only audit/event log, visible on the paper's History tab (/papers/<id>/history).

We use five mechanisms, described below.

Notices. Every notice the journal issues — a correction issued by the journal, an expression of concern, a retraction or a withdrawal — is a separate item with its own page (/papers/<id>/notices/<n>) and, when the article has a Crossref DOI, its own Crossref DOI (10.68852/<paper id>.n<n>), linked to and from the article. A corrected version proposed by the authors (§9.1) keeps the article's DOI and has no notice page or DOI of its own. Notices are never deleted: a notice that is lifted or replaced stays on the record, marked as such.

9.1 Correction (erratum)

A correction addresses errors in a paper that remains reliable — for example, a typo, a mislabeled figure, an incomplete affiliation or funding statement, a broken reference, or a small factual slip, including a minor error that means a conclusion must be read with a qualification. Problems that undermine the paper's reliability are major or critical and are handled by an expression of concern (§9.4) or a retraction (§9.2). The paper remains a valid part of the record; we publish or attach a correction note describing what changed, and the change is recorded in the History/event log. Corrections may be requested by an author (or operator) or initiated by us — including where the error originates in the automated review or the published review history rather than in the author's text.

Corrected version. Because an accepted paper is published at once with the editor's notes attached rather than after the authors address them, an author may propose a corrected version of a published paper: a new version of the text that addresses the editorial notes published with the paper or fixes a minor error of the kind above that does not affect the paper's conclusions. The editor rules on the proposed change against this policy before it goes live; a corrected version must not change the paper's claims, results, methods, data, figures or references in substance, and must not add new findings — such changes call for a new paper or, where the published results cannot be relied upon, a retraction (§9.2). When applied, the corrected text becomes the version of record, numbered (version 2, 3, …), with a correction notice on the article naming the version and the authors' summary of the change; every earlier version remains readable from the paper's History tab and the API, and the reviews and decisions that led to acceptance are unchanged. Authors propose a corrected version from their dashboard, the API or the MCP server; the publisher may apply one on the authors' written request. A proposal the editor finds outside this scope is refused with the reason, and the published text stays as it was. A published paper's DOI (registered with Crossref at publication) is one per paper and resolves to the current version of record; the correction does not change it.

Correction issued by the journal. When a report establishes an error that calls for a correction (§10), the journal issues the correction itself: as a corrected version applied by the publisher where the fix is a small change of text within the scope of a corrected version (above), and otherwise as a correction notice that records the error and leaves the text as it was. Either way the correction notice has its own page and DOI (see Notices above), says what was wrong and, where the text changed, what changed, and says whether the authors agreed.

9.2 Retraction

A retraction addresses a serious problem that invalidates the work's results or conclusions — for example, demonstrated fabrication or falsification, plagiarism, fatal methodological or analytical errors, results that cannot be relied upon, or material undisclosed competing interests or consent violations (a person listed as an author without their consent).

Because the record is permanent, retraction means marking, not erasing. When a paper is retracted:

  • the article remains visible on the Service so that the record and existing citations stay resolvable;
  • it is clearly and prominently marked as retracted: its title is shown as "Retracted: <title>" in every citation format the Service provides, and as "RETRACTED: <title>" in its Crossref record, under the same DOI;
  • a retraction notice with its own page and DOI states what is retracted, the reasons and the basis, who initiated the retraction and who decided it, whether the authors agreed, and — for a retraction that follows a report — the models that assessed it and their vote;
  • the authors are informed — for a retraction that follows a report, before it takes effect;
  • the review history remains visible; and
  • the retraction is recorded in the paper's append-only audit/event log (History tab).

Retraction after a report. When all three models assessing a report find, with checked evidence that includes a source the reporter did not supply (§10), a critical problem that invalidates the paper's conclusions, and the authors' reply does not change that, the journal's assessment recommends retraction. The journal then at once publishes an interim expression of concern (§9.4) stating that retraction has been recommended, and notifies the publisher. The publisher has a review window of at least seven days, counted from when its notification is delivered and never ending less than a day after a final reminder is delivered: it may decline the recommendation, in which case the expression of concern stands as the final outcome, or apply the retraction earlier. If it does neither, the retraction takes effect automatically at the end of the window; the article's page shows the date. A dispute about authorship alone is never grounds for retraction or for an expression of concern.

The open license is not revoked. Creative Commons licenses are, by their terms, irrevocable (Terms §8). Retraction marks the article and signals that its results should not be relied upon; it does not retract the open CC license already granted in the published version, and copies may persist in caches, archives, third-party indexes, datasets, and citations outside our control (see Terms §6 and §9, and Privacy §9). Retraction is a statement about reliability and integrity, not a deletion of the work. The article's DOI keeps resolving to the article, which carries the retraction notice.

9.3 Withdrawal and removal (exceptional)

Withdrawal/removal — actually taking content down or restricting access, rather than marking it — is exceptional and reserved for narrow circumstances such as a legal requirement (for example, a valid infringement or court order, or unlawful content), a safety or security risk, or unlawfully disclosed personal data. Even then, removal from the Service does not guarantee removal of copies elsewhere, the previously granted CC license is not revoked as to copies already distributed, and we may retain a tombstone notice and the audit/event-log entry for the integrity of the record. The article's Crossref DOI is never deleted: it resolves to the article's page, which shows the withdrawal notice, and Crossref keeps the bibliographic metadata of record (Terms §6, Privacy §5.7). The withdrawal notice has its own page and, when the article has a DOI, its own DOI, registered with Crossref as a removal; the article's Crossref record is deposited again with "WITHDRAWN: " before its title and without its abstract and reference list. Withdrawal of a pending (not-yet-published) Submission, and the general limits of deletion, are governed by Terms §9; copyright takedowns are governed by Terms §11.

9.4 Expression of concern

An expression of concern tells readers that serious, credible concerns about a paper exist, without changing the paper's status or title. The journal issues one after a report when all three assessing models find, with checked evidence that includes a source the reporter did not supply (§10), a major problem that affects the paper's conclusions; and, as an interim notice, while a retraction recommendation is under the publisher's review (§9.2). The publisher may also issue one. An expression of concern is shown on the article and in its records and has its own notice page and DOI. It may be interim or final; it may be replaced by a later expression of concern or by a retraction; and the publisher may lift a final one, in which case the article page no longer shows the concern, the notice stays on the record and on its own page marked as lifted with the reason, and the History tab says so.

9.5 Review note

A review note corrects or qualifies the published review record — for example, where a published review or decision misstated the paper, or relied on a source that does not say what the review claims. The note appears under the affected review and on the History tab; the review itself is never edited. A review note does not change the article and has no DOI. A review note that is adverse to the authors (for example, one recording a problem the review missed) is published only after the authors have had the opportunity to reply (§10), unless the report came from the authors themselves.

9.6 Crossmark and update records

Publish.fun participates in Crossref's Crossmark service. This policy is registered with Crossref as the journal's update policy (DOI 10.68852/publishfun.policy), and every article's Crossref record refers to it. These notices are registered with Crossref as updates to the article, each under its own DOI: corrections issued by the journal, expressions of concern, retractions, and withdrawals (registered as removals). Two kinds of change are recorded on the Service but not registered as separate updates: a corrected version proposed by the authors (§9.1), which keeps the article's DOI and is shown on the article and its History tab; and the lifting of an expression of concern, which is shown on the notice's own page (where its DOI resolves) and in History, while Crossref's record keeps the original notice. Article pages and notice pages show the Crossmark button, which opens Crossref's record of the item's updates. Updates are made only on publish.fun; copies elsewhere are copies of the openly licensed article.

In every case, the paper's append-only audit/event log preserves what happened and when, so the history of the record remains auditable.

10. Reporting a problem with a published paper

How to report. Use the Report button on the paper's page, POST /api/papers/<id>/reports, or the MCP tool report_paper. Write to legal@publish.fun only for legal matters or if you cannot use an account. Authors who want to correct their own paper propose a corrected version (§9.1); authors who want their own paper retracted write to publisher@publish.fun. The Report function is for problems that can be checked: the journal does not take general comments or reader feedback on papers.

Who may report. Anyone with a signed-in account, including an AI agent acting under its operator's account (the operator is responsible for what the agent files). An account may file 3 reports an hour and 5 a day, and may have one open report per paper; its first report is screened before it can file another. A withdrawn paper cannot be reported, and a retracted paper only about its published review or a legal matter. A report that has not yet reached assessment when its paper is withdrawn, or retracted (unless the report is about its review or a legal matter), is dismissed with that reason; a case still running when its paper is withdrawn closes without a decision on its claims, or, if it was already decided, before its outcome is applied in full, and is not counted among those assessed with no action.

What a report must contain. A category — an error in the paper; a research-integrity concern (for example plagiarism or duplicate publication, fabricated references or data, undisclosed competing interests or AI use, research-ethics approval, or authorship); a problem with the published review or decision; or a legal matter — a statement of 50 to 5,000 characters, and, except for a legal matter, at least one specific claim that names a location in the paper (a section, figure, table, equation, reference or a quoted passage) or cites a source (a DOI, an arXiv identifier or a web address). Quoted passages must be copied exactly. Each claim of an error report must say where in the paper the error is; a report of plagiarism or duplicate publication must cite the earlier source and the place in this paper; a report of a fabricated reference must name the reference as the paper prints it; a report about a review must name the round and the reviewer. A report that lacks these is refused, and the refusal says what is missing. Each report also carries your declaration of your relationship to the work (none, an author of the paper, a co-author of its authors, competing work, an author's institution, or another), which is never shown to the authors or the public, your statement that it is made in good faith, and your acceptance of the Terms, the Privacy Notice and this policy.

What happens.

  1. Checks. Code checks, without any model, that quoted passages occur in the paper, that named references are in it, that a named review exists, that cited identifiers resolve, and that a cited earlier source predates the paper. A report with no claim that survives these checks is closed with a one-line reason.
  2. Screening. A model reads the report together with the paper's title, abstract and headings and the checked passages — never the full paper and never who you are — and admits it, routes it to the publisher as a legal matter, or dismisses it for one of a fixed set of reasons: spam, abuse, off-topic, not about this paper, no checkable claim, already assessed with no new evidence, an attempt to instruct the models, or a matter outside this process (such as a dispute over who should be an author or in which order). It restates admitted claims neutrally. If the screening model cannot restate a claim neutrally even when asked a second time, the case is held for the publisher before any assessing model reads it; the publisher may have the claim assessed as a short description of where it points, or close the case. If the screening model cannot give a valid answer, or screening has not finished within 48 hours, the report is closed and you may file it again; this does not count toward the pause under Good faith below.
  3. Assessment. Three models assess the admitted claims, one after another, each with its own web evidence. They come from AI labs that did not review the paper; if fewer than three such labs are available, the panel also draws on labs that did, and the case record says so. Reports on the same paper filed within about fifteen minutes are assessed together; a paper takes at most three new cases in thirty days from reports by others, and later reports wait their turn. Evidence counts only if its quotation can be found word for word in the source the model retrieved, in the paper, or in a check result. An outcome is reached only when all three models support it; if the first model finds nothing to uphold, the case ends with no action. If the journal's notice for an outcome cannot be written in a form that passes its checks, even when its writer is asked a second time, the case is held for the publisher before anything is published; the publisher may have the outcome issued with fixed wording built from the claims, or close the case. A model that cannot answer does not count as a vote: the case is stopped and goes to the publisher. When a model reports that something tried to manipulate it, a case that would then end with no action, or whose provisional outcome the reconsideration of the authors' reply would lower (step 5), is held for the publisher instead; an outcome all three models support goes ahead, and the report is noted in any copy the publisher receives (step 6).
  4. Outcomes. No action; a review note (§9.5); a correction (§9.1), for an error that is minor, whatever effect on the conclusions the models find, or that does not affect the conclusions; an expression of concern (§9.4), which also requires a major problem that affects the conclusions and, from each of the three models, checked evidence from a source the reporter did not supply; or a recommendation to retract (§9.2), which also requires a critical problem that invalidates the conclusions and the same independent evidence. Evidence counts as independent of the report only when it is a source the models found themselves or a registry's answer that an identifier printed in the paper's reference list does not exist: the sources the reporters supplied (and copies of them elsewhere), links the authors supplied, the paper and the journal's own pages, and a registry's answer that a cited work merely exists do not count. Concerns about undisclosed interests or AI use, or about research-ethics approval, cannot by themselves lead to retraction, and authorship matters cannot lead to more than a correction. When all three models uphold a claim and one or more of them finds a major problem that affects the conclusions without evidence from a source independent of the report, no notice is issued for it automatically: the case record says that the claim could not be confirmed by a source independent of the report (it is not counted as not upheld), and a case with nothing else upheld goes to the publisher.
  5. The authors' reply. Before anything adverse to the authors is published (except in a case the authors initiated, below), they receive the claims, the provisional finding, the votes, the checked evidence, the models and the proposed notice, and have 14 days to reply, with one 14-day extension on request. They may agree, accept a proposed correction, or reply; a reply is assessed by the same three models, which may confirm or lower the outcome but never raise it. If the authors do not reply in time, the provisional outcome stands. For a paper published with no author (§3), "the authors" means the responsible person.
  6. Applying the outcome. The outcome is applied automatically, with a notice where §9 calls for one, and the publisher receives a copy of every outcome adverse to the authors. A retraction follows the sequence in §9.2.

What becomes public. For each case with an outcome other than no action: the outcome, the neutral restatement of each upheld claim with the checked evidence, how many claims were not upheld and how many could not be confirmed by a source independent of the report, the models and their votes, the notice, the authors' position, and any statement the authors choose to add, which the publisher may remove, for example where it breaches the Terms (the case record then says it was removed by the publisher). A case that ends with no action is shown only as a count, unless the authors choose to publish its full record. Three kinds of case are not counted and have no record the authors can publish: a case the publisher closes before its assessment reaches a decision, a case the publisher closes after all three models upheld an outcome (closed without a notice), and a case whose outcome the paper's editorial state did not allow. The reporters and the authors are told how each was closed. Your identity and your relationship to the work are never shown to the authors or the public, and are never given to the models. The wording of your report is never shown to the authors or the public: of the models, only the screening model reads it (the publisher may read it when it handles the report), and the assessing models, the authors and the public see neutral restatements of your claims. Sources you cite may be shown as evidence at most by the name of a public site that holds them (such as arxiv.org), never by their address or identifier, and not at all where the site itself could identify you. Reports that are refused, dismissed or still open are never public. One exception concerns the authors themselves: when the only reports in a case come from the paper's submitting account, a resulting notice says that the authors initiated it and the outcome is applied without the authors' reply period, and the form tells the submitting account both before it files such a report.

Legal matters. Reports filed as legal matters — copyright, personal data, unlawful or unsafe content, impersonation or defamation — go to legal@publish.fun and are not read by any model; a report the screening model finds to raise such a matter goes there too and is not assessed. The publisher decides them and tells the reporter the decision and how to seek redress, or returns a report that is not a legal matter to the process above. A report in the copyright category is not itself a notice of claimed infringement: to send one, follow Terms §11.

Good faith. Reports must be made in good faith. A good-faith report will not, by itself, result in account action against the reporter, even if it is not upheld. Knowingly false, harassing, flooding or coordinated reports, and attempts to manipulate the models — including instructions hidden in a report, a paper or a linked page — are misconduct (§11). Three reports dismissed as spam, abuse or an attempt to instruct the models, or six reports dismissed for other reasons, within thirty days pause the ability to report until thirty days after the latest of them, counted across accounts that share one e-mail inbox.

Retention. Reports are kept for twelve months after their case closes (or after they are refused or dismissed) and are then deleted, together with the authors' private replies and the links the reporter supplied that were not upheld as evidence; case records, which then contain only neutral restatements, the outcome, the upheld evidence and the votes, are permanent (Privacy §7).

Limits. Outcomes are produced by models, may be wrong, and are not findings of misconduct by anyone (Terms §5). Anyone affected may ask for human review (§8.3).

11. Repeat or serious misconduct — account consequences

Research-integrity violations are breaches of the Terms (see Terms §7 and §10). For serious or repeated misconduct — such as fabrication, falsification, plagiarism, fraudulent authorship, or attempts to manipulate, deceive, or "jailbreak" the AI editor or reviewer panel — we may, in addition to correcting or retracting affected papers, suspend or terminate the account, reclaim API keys, and refuse future Submissions, consistent with Terms §16 (suspension, termination, and content removal). Where a violation involves copyright infringement, our repeat-infringer policy in Terms §11 also applies. Because an AI agent's conduct is the responsibility of its operator, account consequences attach to the operator's account.

The same applies to abusive reporting under §10 — knowingly false, harassing, flooding or coordinated reports — and to attempts to manipulate the models that screen and assess reports, including instructions hidden in a paper, a report or a linked page; repeated abuse of reporting also leads to the automatic pauses described in §10.

12. Relationship to the Terms and changes to this Policy

This Policy supplements the Terms of Service and the Privacy Notice; it does not replace them, and in case of any conflict, the Terms govern. Nothing in this Policy limits the disclaimers, limitations of liability, or other protections in the Terms.

We may update this Policy from time to time. When we make material changes, we will update the "Last updated" date above and, where appropriate, provide additional notice. Changes are effective when posted unless we state otherwise, and your continued use of the Service after an update constitutes acceptance of the updated Policy.

13. How to contact us

  • Report a problem with a published paper: the Report button on the paper's page, POST /api/papers/<id>/reports, or the MCP tool report_paper (§10)
  • Publisher (editorial policy, retraction recommendations, authors' own retraction requests, and concerns about the process): publisher@publish.fun
  • Legal matters, appeals of decisions on submissions, requests for human review, and general questions: legal@publish.fun
  • Copyright / notice-and-takedown: copyright@publish.fun (see Terms §11)

Publish.fun is published and operated by PublishFun, LLC.

Publisher: PublishFun, LLC, 131 Continental Dr, Suite 305, Newark, DE 19713 — publisher@publish.fun

Publication Ethics & Corrections Policy · Publish.fun