Verification becomes the measurable bottleneck, from math proofs to office work

I'd call this space competitive rather than crowded, and the two ideas are in different shape. Generic referee triage for publishers is already largely taken. MDPI runs AI integrity screening on every submission, Springer Nature has Geppetto, Proofig handles images, platform vendors like Silverchair sell gatekeeper AI, and Google and OpenAI give authors free correctness checkers. A startup selling AI-origin detection plus triage would be up against incumbent tools and the publishers' own in-house teams, and AI detection itself has well-known accuracy problems. None of the publisher tools I found judge novelty, though. Formal proof certification is less settled. Math Inc, Harmonic and Axiom have the strongest autoformalization technology, but they present themselves as proof generators and research labs. I found no product that takes a submitted manuscript and returns a neutral pass, fail or gap certificate that a journal or funder can rely on. The AXLE example, where Axiom's engine checked proofs written by other labs' systems, hints at that role, but none of them sell it as a service. So the opening is to be the independent auditor rather than another prover. That means a trusted third party that does not generate proofs (so it has no conflict of interest when the labs publish results), checks that formal statements actually match the paper's claims (the known weak point, since Lean only verifies the theorem as stated), and fits into editorial systems with standard gap reports. The main risks are that the prover labs add certification themselves or partner directly with journals, that open-source document-level autoformalization from EPFL and LeanAide makes the pipeline a commodity, and that most research math still can't be formalized cheaply because Mathlib's coverage is limited. A realistic start is narrow: certifying AI labs' own headline claims and serving specific combinatorics or number theory journals, before trying to sell broad triage to publishers.

• Math Inc (Gauss / OpenGauss)

  Venture-backed autoformalization agent. It formalized the strong Prime Number Theorem in Lean in about three weeks and helped with the sphere packing formalization. Scientific American reports it can flag and fix minor errors in papers. It is the closest match to the proof certification idea, though I found no per-manuscript certification product or journal deal.

• Harmonic (Aristotle)

  Lean-based prover that reached gold-medal-equivalent performance on IMO 2025. It turns natural-language problems into formal statements and proofs, which puts it in a position to offer verification of claimed proofs.

• Axiom Math (AxiomProver, AXLE)

  AxiomProver autoformalizes and proves research results. In one paper, its AXLE Lean engine served as an independent checker for proofs that Aristotle and OpenGauss generated, which is close to the neutral certifier role.

• EPFL Document-Level Autoformalization (Renaissance Philanthropy)

  Funded academic project building open-source tools that formalize whole mathematical documents, which could commoditize the core pipeline.

• LeanAide

  Open, philanthropy-funded no-code environment that combines AI with Lean for autoformalization.

• Google Paper Assistant Tool (PAT)

  Gemini-based checker for errors in papers. It was piloted on more than 4,700 submissions at STOC and ICML and is positioned for authors to use before submitting. This is a big-company move into informal correctness checking for math and theory conferences.

• OpenAI Prism Paper Review

  Checks whether derivations are correct, whether units are consistent, and whether claims follow from the method. Positioned as a pre-submission rigor check.

• MDPI Ethicality

  In-house AI integrity screening of every MDPI submission: paper mills, AI-generated text, citation manipulation. This is publisher-side triage already running in production.

• Springer Nature Geppetto and Proofig

  Tools publishers already use for AI-text detection and image-integrity checks. They are part of large research-integrity operations that screen millions of submissions.

• Silverchair submission-to-decision tooling

  Vendor of publishing platforms marketing gatekeeper AI that checks whether submissions are complete, compliant and technically sound.

• YesNoError and Black Spatula Project

  A commercial tool and an open-source project that scan published papers for errors at scale (tens of thousands and about 500 papers respectively). Both use informal LLM checking, not formal proofs.

• Igor Rivin arXiv error analysis

  Research system that scanned more than 37,000 arXiv math papers for errors and recommended journal tiers, showing that automated triage of math papers is already feasible.

Checked 2026-10-11