Invigil
Invigil measures how well frontier AI models reason, using formal verification problems written in Lean 4 and marked by a proof checker rather than a human grader. It sells a private, never-published set of such problems to frontier laboratories, government safety institutes and regulated enterprises, refreshed each quarter with material written after the models under test were trained.
Invigil is the first instrument built on the protocol. Each quarterly holdout run is executed under the Antefacts method: the problem set and scoring are pre-registered and hash-committed before the run, the harness is published, the record is hash-chained with independently republished roots, and each run carries an Antefacts certificate that the procedure was followed as registered. Invigil's customers receive a measurement certified by a party other than Invigil. Invigil pays a licence royalty to Antefacts on the revenue of those runs.
Antefacts and Invigil are under common ownership
Antefacts and Invigil are separate companies with separate management, infrastructure, customer contracts and data, owned by the same parent. The licence between them is at arm's length and its terms survive a sale of either company. We state this here because our own rule says a reader of any record we seal is entitled to know it, and because the alternative, that a notifying authority or a losing vendor discovers it, is exactly the failure the method exists to prevent.
What follows from it, in the specification and not merely in policy:
- Every Antefacts pre-registration carries a published relationship field stating whether the subject, or any model inside the subject, is a customer of Invigil or of any other company in our group.
- Where the subject of an evaluation for a certification body or a procurement buyer is itself an Invigil customer, the evaluation proceeds as a published methodology with no scored subject, or not at all. The same rule we apply to our own competitors applies to our group's customers.
- No held-back data, problems, credentials or infrastructure are shared between Antefacts and Invigil. The Invigil problem library is never used as Antefacts held-back material and Antefacts items are never used by Invigil.
- Invigil's customer list is disclosed to Antefacts for the sole purpose of applying the rules above, and is not otherwise used.
- The classification of every relationship is countersigned by the party whose decision the evaluation informs, and misclassification is an admissible ground of challenge open to anyone named.
- The challenge owner for Antefacts evaluations sits outside the group.
None
Antefacts holds no equity in, and receives no fees from, any organisation it has evaluated or expects to evaluate. It has no advisory, implementation or reseller arrangements with any AI vendor. When that changes, this page changes first, before any evaluation the relationship could touch is pre-registered.
Licensed bodies
No certification body is yet licensed. Bodies that license the method will be listed here with the date of their licence and the domains it covers, so that a notifying authority can see who runs the method and since when. Read about the licence for certification bodies.