The AI system carries its own proof.

Most fairness tooling makes a quiet choice on your behalf. It picks one statistical definition of fairness and applies it everywhere. That is convenient, and it is wrong, because fairness is not one question. In lending, the binding constraint may be whether qualified applicants are approved at comparable rates. In hiring, the operative standard may be an adverse impact test that regulators and courts already recognise. In healthcare triage, equal treatment of unequal need is not fairness at all. Pick the wrong measure and you can pass a test while causing exactly the harm the law was written to prevent.

So we do not pick one for everybody, and nobody gets to pick quietly.

Every assessment starts with its context: the industry, the specific use case, the jurisdiction, and the type of decision being automated. From that context the platform queries a knowledge graph to resolve which fairness tests the law and the research literature actually support in that situation. The answer comes back as an ordered list, with the primary test named, each entry traced to the statutes and the peer-reviewed work that stand behind it.

We stop short of calling that a legal command, because no statute names a formula. The measures also trade off against one another, provably: satisfy one and you can be forced to breach another. Choosing between them is a judgement, and a judgement has to belong to somebody. So the resolved profile is put in front of the system's owner as the recommendation, the owner adopts it or deliberately departs from it, and the seal records both: what was recommended and what was used.

What the owner declares is the case: the sector, the jurisdiction and the kind of decision. What the rule then is, for that case, we resolve ourselves at the moment of issuing, reading our own curation rather than anything sent to us. So an operator can choose which case they are in, and must live with the rule that case carries. Nobody hands us the rule they would prefer to be judged by.

So the seal does not only say that a system passed. It says which test it passed, why that test was the defensible one for that case, and where the grounding came from. The rule is signed rather than asserted, which means the choice of rule is on the record alongside the result.

Enlarge

Context in. Applicable rule out. The rule is frozen into the signed credential, so nobody can quietly change the test after seeing the result.

It is worth being blunt about what this costs us. It means we sometimes hand a paying customer a result they cannot put in a press release, and it means the system's default answer, whenever it is unsure, is no. A seal that always passes is decoration. A seal that can decline is an instrument.

1. 01

   A seal states which rule it enforced, and one of the three is narrow.

   The verdict is decided on every measure the rule marks as binding, not on one. Each is tested against its own bound in its own direction, a measure that only has to be reported can never veto, and a measure the rule requires to hold first blocks the seal outright rather than failing it, because a statistic you cannot interpret is not the same as a system you have caught. Every seal names which of three rules applied. The strongest is resolved by us from the sector and jurisdiction the assessment declares, so nobody can hand us the rule they would like to be judged by. Next is a rule the operator froze at the time of assessment. Narrowest is the single agreed measure, which is what a seal falls back to when no sector rule has been declared, and a seal in that state says so on its face. It is one system, at one moment, and it is not a substitute for a full fairness audit. It does not certify compliance with anything.

2. 02

   Intersections are measured, carried and can fail a seal. What they cannot always do is conclude.

   Every fairness measure runs on combined group keys, so cells such as women over fifty are measured with the same confidence intervals as the headline numbers, and that reading is signed by the measurement engine and carried into the credential. The rule is deliberately asymmetric. Evidence of harm in a cell the study was actually powered to judge, surviving the correction for having tested many cells, fails the verdict outright: a clean result on each attribute separately cannot rescue it, which is the whole point, because that is exactly how the best known cases of this harm were missed. Absence of evidence is treated differently. Cells too small to judge, or suppressed because judging them would expose individuals, cannot fail a seal and instead cap the confidence it may claim, weighted by the share of people they cover, so hiding people in unjudgeable cells destroys the confidence rather than buying a pass. Attempting the analysis is not optional: a seal that carries no crossed-attribute reading is refused rather than issued.

3. 03

   No causal claim, by construction.

   A measured difference is a difference in outcomes, not proof of what caused it. Causal identification requires assumptions our data does not support, so we do not make the claim.

4. 04

   Our verifier is online by design, and it abstains rather than rejects.

   Two of the three checks need no network. Whether the disclosed numbers match what was signed is arithmetic over the credential itself, and the signature verifies against our published key, which anyone can hold in advance. Only revocation truly needs a network, and even that can be answered from a signed status token that stays usable for a bounded window and then refuses rather than going quietly stale. Our own verifier still fetches the key and the status list fresh every time, on purpose: one that trusts a cached key cannot see a rotated one. So checking without a network is a property of how this is built, not something you can do at our page today. And when a check cannot complete, the answer is grey and says exactly that. It is not the red answer, which says the credential may be altered or forged. Not knowing is not the same as catching a forgery, and it is not a pass either.

5. 05

   No live public-sector deployment.

   We have applied to the Canton of Zurich sandbox and are in the matching phase. That is an application in progress, not a deployment, and we will not describe it as one until it is. Our Swiss trust-registry entry sits on the integration, or preview, environment while the federal electronic identity moves towards production, so the rails are real and the registry entry is not yet a production one.

6. 06

   The instrument is further along than the footprint.

   The credential, the issuing service, the resolver and the public verifier all work today, and you can check that yourself in the section above. What we are not going to do is dress that up as adoption. This is early, the deployed count is small, and anyone who tells you otherwise about a system this young is selling you something.