ProofForum keeps source files, AI attestations, human theorem-level reviews and discussion as separate layers. A paper is not treated as mathematically verified merely because an AI check passed, and community agreement is not presented as a Certified Referee signature.
The LaTeX source remains canonical for structure extraction and mathematical metadata. Browser-generated structure data is not trusted as the source of truth. Inline \bibitem entries are also parsed. A citation containing a ProofForum public code such as PF:2026.000001 is resolved to the published paper when possible, becomes clickable, and is shown with that cited paper's current human-review state. Bibliographies stored only in an external .bib file are flagged for manual inspection because that file is not part of the current PDF + single-.tex upload format.
New submissions and new versions remain non-public until accepted. Before accepting a version, the moderation view performs static checks on the stored files: suspicious LaTeX patterns, active-content indicators in the PDF, and an approximate PDF↔LaTeX consistency comparison. High-risk indicators require an explicit override.
The moderation panel also reports reference parsing: number of inline references, unresolved internal ProofForum IDs, duplicate bibliography keys and external-bibliography limitations. These checks are designed for triage and risk reduction. They are not a malware certificate, a guarantee that a citation is bibliographically correct, or a mathematical correctness check. A suspicious or inconsistent result should lead to manual inspection or rejection.
After publication, reviewers work on the mathematical structures extracted from that exact paper version. A review is attached to one theorem, lemma, proposition, corollary, definition, claim or conjecture, not generically to “the paper”.
Role-weighted signals remain distinct. Certified authority is subject-scoped: a Certified Referee or Managing Editor receives that authority only for the mathematical subjects assigned by the administrator (or an explicit All subjects scope). A certified judgment is not silently merged with ordinary community consensus.
Reviewers can select a passage in the PDF and attach one focused comment to that text. The thread stays shallow: replies are allowed, but replies do not branch into unlimited nested discussions.
Comments can be edited by their author; issue threads can be resolved or reopened by the permitted participants. Comments do not turn a theorem green: formal human-verification state and discussion metadata remain separate.
Can: view public papers, add human reviews after email verification, comment on published versions, reply to discussions, and record an LLM check after reviewing that exact version when permitted by the site rules.
Cannot: issue the Certified-Referee-only “incorrect” verdict or the purple expert annotation. Submission may also require institutional eligibility.
Can: participate with a verified institutional identity, submit positive/concern reviews and use the normal discussion tools. A valid institutional verification is checked when this role reviews.
Cannot: make the Certified-Referee-only incorrect judgment or purple expert annotation solely by virtue of institutional verification.
Can: submit certified positive, concern and incorrect judgments, use the purple “interesting proof argument” annotation and exercise other certified-review powers inside the subject scope assigned by the administrator. The scope may contain one or several subjects, or the explicit All subjects tag.
Cannot: exercise certified authority outside that scope, make another person’s review disappear, or convert community consensus into a certified signature. Each reviewer’s record remains attributable.
Can: act as the subject-level editor for every version of papers whose primary/secondary/additional subjects intersect the assigned scope. That includes inspecting draft, pending, published, rejected and withdrawn versions; running safety/reference checks; adding AI-check attestations; accepting, rejecting or unpublishing versions; and exercising Certified-Referee-level theorem-review authority inside the same scope.
Cannot: manage users, assign editor/referee privileges, administer referee applications or institution verification, permanently delete repository records, or override a high-risk security block. Those remain site-administrator powers. Only the administrator assigns/removes Managing Editor status and its subject scope.
Can: see their own pending/rejected versions, upload later versions while eligible, add AI-check attestations, participate in discussion, and resolve issue threads on their paper.
Cannot: publish a pending version by themselves or bypass server-side eligibility, moderation and file-access checks.
Every uploaded revision receives its own version record and its own recorded license. A published version keeps that license permanently; a later version may use another permitted license if the submitter is authorized to grant it. Reviews, comments, AI checks, extracted structures and file checks refer to the correct version. The Versions view can compare revisions without merging the evidence from two different versions.
A later version can address an earlier blocking comment, but the historical comment remains part of the earlier version’s record. This preserves provenance instead of rewriting the past.
Reading papers, submitting work, reviewing structures and applying for Certified Referee status are intended to remain free. The site may still have unavoidable operating, security, compliance and staffing costs, but those costs must not be converted into pay-to-publish, pay-to-certify or pay-for-a-favorable-review mechanisms.
The purpose is practical as well as ethical: separating money from acceptance and referee judgment reduces incentives for favoritism, pressure and corruption. No payment should change a mathematical verdict, suppress a concern, accelerate certification, or buy extra visibility.
ProofForum is designed to make mathematical provenance and review evidence inspectable. It does not claim that an AI check proves correctness, that a static PDF/LaTeX scan guarantees safety, or that a high community score replaces a rigorous proof. The value of the site comes from keeping these signals explicit and attributable.