An open repository for AI-generated mathematics.

Reuse before regeneration

ProofForum exists partly to reduce duplicated AI work. If a theorem, proof strategy or closely related result has already been generated, checked and discussed, researchers should be able to find and reuse that record instead of sending millions of near-identical prompts to multiple models. Better retrieval can save time, compute and repeated model inference, and can therefore reduce the environmental cost associated with unnecessary AI usage.

Against AI slop

ProofForum is explicitly against high-volume, low-effort AI slop. Uploaders are responsible for basic screening before publication, including a complete-paper AI check. Repeated submission of incorrect, meaningless, duplicated or negligently unchecked material can lead to rate limits, temporary suspension or account removal.

For this prototype, the submission policy is at most five new papers per rolling seven-day period per account. In production this limit should be configurable and reputation-sensitive. A pattern of structures marked incorrect by certified referees, or repeated unresolved concerns across an uploader's papers, should trigger moderation review and can suspend further submissions.

Institution-verified accounts

Submission accounts should be tied to a real academic or research identity where possible. ProofForum can support institutional verification through a verified institutional email challenge and, preferably, university or research-organization single sign-on (SAML/OIDC). A domain name alone is not treated as proof of affiliation.

In this prototype the verification state is represented explicitly on the account and enforced for server-side submission. Production deployment should connect that state to a real verification workflow and retain an auditable record of when and how affiliation was confirmed.

Who gets credit?

The primary provenance identity is the person who generated and submitted the paper. ProofForum therefore uses Uploaded by rather than pretending that a conventional authorship model always describes AI-generated work accurately. The uploader is responsible for the prompt-driven submission and for placing that version in the public record.

Certified referees sign their work publicly

Certified verification on ProofForum is not anonymous. When a certified referee verifies a theorem, lemma, proposition or other primary structure, that verification shows the referee's name together with the scope and version reviewed. The point is accountability at the place where the mathematical claim is certified.

ProofForum does not publish a browseable directory of certified referees. Identity is disclosed through signed verification events, not through a public roster. Comments and informal discussion remain separate from certification: a comment does not turn a structure green.

What the colors mean

Green means strong human support: either a certified ProofForum referee has signed the exact structure, or at least five independent academic referees with verified institutional email have positively checked it. Yellow means positive community support that has not reached that institutional threshold. Checks from normal users remain yellow regardless of quantity; one to four academic-referee checks are also yellow. Orange means a reviewer has raised a substantive concern or requested revision. Red is reserved for a structure judged incorrect or rejected after human review. Blue is AI/model evidence only. Unreviewed structures remain neutral gray.

The primary units are structures extracted from LaTeX: theorems, lemmas, propositions, corollaries, claims, conjectures and definitions. Readers can then open a structure and discuss a specific phrase, equation or step inside it.

Minimum admission rule

Every uploaded paper must include its PDF and LaTeX source, come from an institution-verified active account, and carry at least one full-paper AI check attested by the uploader. The record states the provider, exact model, result and that the model was asked to inspect the complete paper. This is a minimum provenance requirement, not a substitute for human proof checking.

Versions and control

Versions are immutable records. A new version does not overwrite the old one: ProofForum keeps prior source and a visible diff. Only the original uploader, a certified ProofForum referee, or an administrator may publish a new version of an existing paper. Ordinary users may comment, but they cannot modify someone else's submission.

Scope

• AI-generated or substantially AI-assisted mathematical manuscripts with inspectable LaTeX.
• Explicit model provenance and at least one complete-paper AI check.
• Human verification attached to mathematical structures, with certified referee signatures and visible institutional-review consensus.
• Open discussion and versioned correction without erasing earlier states.

Free access and independence

Reading, submitting, reviewing and applying for referee certification are intended to remain free of charge. ProofForum may need to cover unavoidable hosting, security, compliance and staff costs, but payment must never buy acceptance, a favorable review, referee certification, ranking, visibility or the resolution of a mathematical dispute.

This separation is deliberate: keeping mathematical judgment independent from payments reduces financial pressure on authors and reviewers and helps limit conflicts of interest, favoritism and corruption risks. The long-term aim is to preserve this principle as the project grows.