Method knowledge is not left to the model’s training: it is a curated, versioned canon inside the product, and each audit finding declares which canon module it is grounded in.
The problem
A generic model’s answer blends whatever training saw: vendor blogs, old forums and the manual that suits it, without saying which one weighed in. Auditing that answer is impossible, because there is no way to know where it came from.
The report is grounded in a curated canon: DAMA-DMBOK, the modelling of Kimball, Inmon and Data Vault, Data Mesh, TOGAF, Well-Architected and more, versioned in the product. The finding cites the module, and the citation can be checked.
How it works
The audit report examines the material you sent and returns a verdict: approved, approved with reservations, or rejected. Each finding carries area, evidence in your own artifact, impact and recommendation.
And it carries the line usually missing from AI reviews: the grounding in the canon, pointing at the methodology module that supports the critique.
Audit report: transactional e-commerce model
Evidence: stock_movements has no materialization; the balance is summed on every read.
Impact: above 1 million movements, the balance query becomes the checkout bottleneck.
Recommendation: materialize the balance, with periodic reconciliation against the sum.
Grounded in the canon: modeling-kimball-inmon-datavault
Every finding in the report becomes a project attention point with its own state, so it does not die inside the document.
Each finding in the report becomes a project attention point with its own state. The document may be forgotten; the attention point stays open until someone decides on it.
The deliverable’s seal shows what was verified in that document, and what was not shows as not measured, never as approved by omission.
Transactional model for e-commerce
Version 2 · audited · 6 entities, 31 columns
| DDL executed on a real database | done |
| Business questions answered by query | done |
| Adversarial review | done |
| Proof under load | not measured |
Second opinion, from a different model
Disagrees with normalizing stock_movements: at high volume, a balance computed by summation becomes the bottleneck. Suggests materializing it. The disagreement was recorded, not resolved.
Under the hood
Versioned methodology modules, applied as the subject demands.
What comes next
An adversarial reviewer attacks through cost, risk and simplicity; answers get verdicts, and an objection that stands is kept in the document.
See it from the inside →ProofMachine verificationScripts executed on a disposable PostgreSQL, business questions answered by query, and the seal showing what was and was not verified.
See it from the inside →As-isA real as-isDiscovery happens through guided investigations: read-only scripts you execute, with the result coming back tied to the question.
See it from the inside →Paste the structure, get the most serious findings in about two minutes, and decide later whether an engagement is worth opening. No account, no card.