Certify, don't vote: machine-checkable certificates as the arbiter of multi-model critique
Multi-model review is now a common way to check work done with language models: several model families critique a result, and their agreement is read as confidence. We describe a method that takes agreement out of the decision. Model families generate hypotheses and objections. A claim is promoted when a mechanical arbiter accepts it (a DRAT proof checker, a Lean kernel, or re-simulation of an explicit circuit), or, through one declared fallback, when a hand-proved ordering lemma has been checked by three model families and the claim says so. Two further exceptions are declared where they occur: a lifting theorem that is machine-checked in Lean except for one premise stated in prose, corroborated by DRAT at small n, and Lean files whose proofs also trust the compiler. The method has three parts. A certification ladder states which certificate each class of claim needs and what a fallback must declare. A dated, additive claim ledger is guarded by a lint that requires a certificate reference or one of two declared alternatives. Since 2026-09-12 reviewers asked to judge a claim are asked for rankings and discriminating tests, and a verdict label they still write is recorded as data, not as a decision; since 2026-10-11 the written protocol says so (Section 5.3). Reviews of a whole manuscript before deposit still return findings, and the verdict on each finding comes from a verifier that is itself a model (Section 5.5). We report the method as one research program in exact circuit synthesis used it, and that program is the worked case. The case shows why a vote cannot be the arbiter. One model family proposed a gate-ordering constraint and a second confirmed it as sound. A third found a concrete circuit that the constraint wrongly excluded. The case also tested the method against a mathematical result rather than a benchmark. The program refuted two theorems of a published preprint: the unit-gap bound, and the classification of unit-gap circuits that the worked case is about. The author withdrew the preprint, and the withdrawal comment cites the program's counterexample to the first theorem and its certificates. Only level 1 of the ladder, plain DRAT without the added gate-ordering constraint, is implemented. Level 2, VeriPB dominance certificates, is a design: its checker is built, and there is no instance encoding, no working proof-logging solver and no dominance derivation. Our own planning notes described the nightly durability check as covering the certificate corpus. It covered one certificate. Durability today is a nightly re-verification of 32 certificates: the one the job always covered and the 31 shipped with the public bundle, added to the job on 2026-10-08. The job's log for the 31 shows all of them passing on each of the 44 nights from 2026-10-09 to 2026-11-21. Checked once on one laptop, the 31 passed in 21 s of checking (46 s for the whole run), and the checker rejected three deliberately broken proofs. The panel's process records give counts of calls and exit codes. They do not give a true-positive rate, and we do not report one. Evidence. The evidence files cited by the paper are in two public records. doi:10.5281/zenodo.24261805 holds the re-verification run and the other evidence files from the author's companion repository. doi:10.5281/zenodo.24421990 holds the cited files of the research repository, in English, and the source of this paper with every evidence marker.
Authors
- Luiz Antonio Busnello (ORCID: https://orcid.org/0009-0007-5911-8141)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-21
- DOI
- https://doi.org/10.5281/zenodo.24422003
- Primary Topic
- Formal Methods in Verification
- Type
- preprint