Commitment before assessment
Criteria, weights and pass thresholds are published and hash-committed before any subject is assessed. Nothing moves once the data arrives.
The problem
Attestation, not evidence
A corporate picks a supplier after a bake-off whose criteria were never published. A standards body names a preferred auditor after a workshop. A regulator is asked to accept a conformity assessment whose underlying test cannot be rerun. In each case the decision may well be right. There is simply no way for anyone outside the room to tell.
Antefacts builds the layer underneath, so that the working can be shown.
Assurance model
Two kinds of claim, never mixed
Most of this market sells one blurred assurance. We separate the two claims that can honestly be made, and we are explicit that only one of them admits proof.
Claim tiers
| Axis | What is claimed | How | Tier |
|---|---|---|---|
| Process integrity | This evaluation was conducted in accordance with its own pre-registration, and its record is unaltered | Machine-checked, on every evaluation | Proved |
| Subject performance | The system measured at a stated value, within a stated interval, under a stated configuration | Pre-registered statistics, or pre-registered scoring | Bounded Registered Never proved |
Set out in full on the method page. No evaluation we run will assert that a system is proved safe, proved correct, certified or compliant. That restraint is not modesty. It is the only position that survives a technical assessor reading the method closely, and it is what makes the first claim worth anything.
Construction
Cryptographic, and deliberately not a blockchain
Every record is hash-linked, signed and timestamped by an authority we do not control, and the chain roots are published to independent locations so that we cannot quietly rewrite our own history. That is a hash chain in a database with independent republication: the boring, auditable construction. A distributed ledger would add nothing to the guarantee and a great deal to the doubt, so our specification prohibits one.
Mechanisms
What that changes
01
02
Held-back inputs
Subjects receive the requirement specification. They do not receive the scoring inputs in advance, and they are barred from retaining them.
03
Rerunnable by anyone
The scoring harness is published. A third party can obtain it, rerun the evaluation and check every hash without contacting us.
04
A record that resists rewriting
Append-only, hash-chained, independently timestamped, with roots republished where we do not control them.
05
A route to contest
Every participant may obtain the harness and rerun its own submission. A challenge is a defined record with a deadline and an outcome, and it enters the ledger whichever way it is decided.
06
Deviations on the record
Evaluations depart from plan. When ours do, the departure is recorded and published, including where the results themselves stay confidential.
What we offer
Two ways to use the method
Licence
For certification bodies
An annual licence to the method and the platform, co-branded. Your assessors run the evaluations, you issue the opinion, the evidence carries both names and survives an appeal. Built for the bodies that will assess high-risk AI under the EU AI Act from 2 December 2027.
Service
For procurement
A fixed-price evaluation of the two to four systems you are choosing between, on criteria you author and we seal before any vendor is seen. The losing vendors can rerun it. That is why they do not dispute it.
In production
The first instrument built on the protocol
Invigil, which measures the formal reasoning of frontier AI models on Lean 4 problems they have never seen, runs each quarterly holdout under this method. Its customers receive a measurement certified by a party other than Invigil. Invigil is under common ownership with Antefacts; what that means for our neutrality, and the rules that follow from it, are set out on the licensees and disclosures page.
Timing
Fifteen months
From 2 December 2027 a notified body's assessment of a high-risk AI system is a legal act. Providers appeal it, competitors contest it, notifying authorities review it. Bodies must have chosen a method before then, and bodies choose once. The timetable is on the certification bodies page and our reading of the regulation is in the notes.
Audience
Who this is for
Certification and notified bodies that need a defensible technical method for conformity assessment and have no intention of building one. Corporate procurement functions whose supplier selections have to survive challenge from the vendors that lost. Anyone whose signature ends up on a decision they cannot currently show the working for.
Release
What we publish, and when
The protocol specification, the scoring harness and the machine-checked development are given away. They are released as one bundle together with the first conforming evaluation, because a specification published alone is a document, and the four published together are a demonstration.
Already here: how the method works, what it does not buy, and the rules we hold ourselves to.
Etymology
The name
From the Latin ante factum, before the fact. An antefact is a thing fixed before the event it governs: the criteria, the weights and the thresholds, sealed before any data exists. That is the whole method in one word.