On this page
Proof artifact registry for research teams
Edition: 8 October 2026. Initial decision: Prototype now. The product interpretation and pricing are hypotheses; independent proof and buyer validation remain open.
Research finding
The collection combines manuscripts, separate challenge statements, solution modules, checker configurations, and different coverage levels. The 416-configuration static preflight found all referenced modules and two configurations enabling an external kernel; these are metadata observations rather than proof verdicts.
Problem and buyer
An AI research lab or university formalization group must establish exactly which statement was checked, with which toolchain, assumptions, and artifact version. Ad hoc links and successful builds leave this review work scattered across logs and repositories.
What the finding could enable
A registry can package a reproducible verification record, attach it to a manuscript claim, and expose an API for downstream users. The new collection provides a substantial integration corpus. The business value is evidence management and reproducibility; proof registries were possible before this release.
Technical and commercial limits
The prototype only parses metadata. Actual verification needs trusted challenge definitions, controlled imports, compatible Lean tooling, a correctly configured sandbox, and preserved checker logs. Passing a checker does not establish practical usefulness or the intended interpretation of a theorem.
Minimal architecture
Git ingestion -> content hashes -> claim and statement records -> isolated verification jobs -> signed run metadata -> team review dashboard. Begin with the existing proof_preflight.py scanner; add job isolation and full logs before producing any verified badge.
Existing alternatives and differentiation
Comparator is an existing open-source checker. Compete on reproducibility, claim-level indexing, team review and managed operations; do not rebuild its kernel checks as a proprietary differentiator.
Monetization hypothesis
Hypothesis: AUD 2,000-5,000 for a bounded onboarding and reproducibility audit, then AUD 300-1,500 per team per month plus transparently billed compute. The specialist audience may be too small for a standalone company; integration services can test willingness to pay first.
Validation experiment
Reproduce ten selected checks in an appropriate environment and have a reviewer reconstruct every conclusion from the saved record. Target zero unsupported verified labels and a material reduction in reviewer preparation time relative to manual assembly.
Conditions to reject or defer
Defer if artifact maintainers already solve the workflow with CI at negligible maintenance cost, or if buyers will not pay for reproducibility rather than theorem discovery.
Next concrete action
Specify the verification-run schema and build a local record viewer using the static report; investigate a suitable Linux checker environment without provisioning paid infrastructure.
Source evidence
Repository sources are pinned to revision fd4aeeb2ee4fc729c18d98444fed42fd0529eeeb. Selected scope notes and manuscript statements have been reviewed to the extent described above; these links do not represent successful kernel checks.
Family 113
Subject: Approximate counting and entropy of perfect matchings.
- A Fully Polynomial Randomized Approximation Scheme for Perfect Matchings in General Graphs
- Entropy and Face Dimension of the Perfect-Matching Polytope
- Formal scope notes
Family 115
Subject: Sampling and counting contingency tables with arbitrary margins.
- Exact Uniform Sampling of Contingency Tables with Arbitrary Margins
- An FPRAS for Cell-Bounded Contingency Tables
- Formal scope notes
Family 124
Subject: Polynomial-time scheduling on three identical machines.
Family 325
Subject: The complete Crouzeix conjecture.
Current alternative sources
Primary documentation reviewed on 8 October 2026. Product availability demonstrates alternatives, not demand or willingness to pay for this proposal.