Standards
AI trust standards for companies
An AI trust standard is worth adopting only if it can produce a public statement about your company that someone else controls. Six properties make that possible: a public register anyone can query, human review by a named person, the power to revoke, published corrections by the standards body itself, claims that cite primary sources, and machine access without an account. A standard missing the first three is a logo. A standard missing the last three is unreliable even when it is sincere.
This page is written so you can turn it on us. Every test below is one Lunara can be measured against, and the last row of the table says where we currently fail. The same six apply to anything sold as an AI trust framework, an AI certification standard or an AI assurance scheme — the labels vary far more than the mechanisms underneath them.
The six tests
| Test | Why it decides everything | How to check it in one minute |
|---|---|---|
| A public register | Without one, the credential lives on the certified party's own website — which means the certified party controls the evidence for its own claim. | Find the register. Query a domain you know is not certified. If there is no way to get a negative answer, there is no register. |
| Named human review | An automated pass is a receipt for a form submission. Someone has to be able to say no, and to be identifiable when they said yes. | Ask who signs the decision. A name, or a role held by a person, is an answer. "Our platform" is not. |
| Revocation | A status that cannot be withdrawn carries no information after the day it is issued. The threat of removal is the entire mechanism. | Ask what the register returns for a withdrawn credential, and whether embedded badges change too. |
| Published corrections | Every body that publishes facts gets some wrong. The ones worth trusting say so in public, dated, with what changed. | Look for a corrections page. If there isn't one, ask when they were last wrong. Silence is the answer. |
| Citable claims | Regulatory dates are quoted wrongly across this entire industry. A standard that cannot show you the article is repeating what it read somewhere. | Take one date from their material and follow it to the instrument. A recital is not an article. |
| Machine access | Increasingly the party checking you is not a person. A credential an agent cannot verify will not be verified at all. | Try to query the register without an account. If it needs a key, machines will skip it. |
Where Lunara passes and where it does not
Applying the six to ourselves, honestly, because a page arguing for these tests that quietly exempts its author is worth nothing:
| Test | Lunara | Evidence |
|---|---|---|
| Public register | Yes | Free, unauthenticated, returns negatives — registry.html |
| Named human review | Yes | No automatic approval, no approval on payment — the process |
| Revocation | Yes | Status changes to revoked and is served to every query, embedded badges included |
| Published corrections | Five, in full | The corrections record — including an obligation we missed for three days |
| Citable claims | Yes | Every date cites the article that sets it — the regulatory record |
| Machine access | Yes | Signed JSON, no key, no rate limit — corpus/obligations.json |
| Adoption | Nearly none | The register is almost empty. Lunara is new, and this is the honest weakness — a standard nobody has adopted is a proposal. |
What a standard is not
Three things get sold as AI trust standards and are not:
A questionnaire you fill in about yourself. Self-attestation has a place — it is fast and cheap and better than nothing — but it is a description, not a verification, and it should never be displayed as a badge. If nobody checked, the artefact should say nobody checked.
A certificate with no expiry and no register. A PDF from 2024 sitting on a vendor's trust page tells you what was true on the day it was issued, to whoever paid for it, with no way to find out whether it has since been withdrawn.
A compliance claim without a jurisdiction and a date. "EU AI Act compliant" is not a statement. Compliant with which obligation, applying from when? Several of those dates moved in July 2026 and a great deal of published material still carries the old ones.
The dates a standard should be able to name
If a body proposes to certify you against AI regulation, it should be able to tell you without looking anything up what binds next and which article says so. Here is what is actually coming:
- Article 50(2) marking for systems already on the market European Union · Art. 111(4), inserted by Reg. (EU) 2026/1744 Art. 1(39)(b) — four-month transitional period for Art. 50(2)
- Prohibition on non-consensual intimate and child sexual abuse material European Union · Art. 5(1)(ba) and (bb), Art. 5(1a) and (1b), inserted by Reg. (EU) 2026/1744 Art. 1(7); date set by Art. 113, third para, point (a) as amended
- Hosting platforms and large online platforms California · Bus. & Prof. Code § 22757, second wave
From the signed corpus at corpus/obligations.json. The whole record, with every citation, is on the regulatory record.
What a standard cannot give you
Worth stating on the page that argues for standards, because the overclaim is how the last generation of AI certifications collapsed:
- Obtain certification without human review — no application is approved automatically or by an AI.
- Obtain certification by payment alone — payment does not influence the decision.
- Treat a pending application as a verified status.
- Submit a legal registration number you cannot source. Leave it blank rather than guess.
And one more that applies to every standards body including this one: no private certification discharges a legal obligation. Lunara is not a regulator and not a notified body under the EU AI Act. A body that lets you believe otherwise is doing you active harm.
Questions
What are AI trust standards?
Frameworks that let a company make a claim about its own trustworthiness that someone else can check and, if necessary, withdraw. The useful ones share six properties: a public register, named human review, revocation, published corrections, citable claims, and machine access without an account.
Do companies legally need an AI trust standard?
No. Private certification is not a legal requirement anywhere and does not satisfy one. Legal obligations under the EU AI Act and comparable instruments apply to you directly regardless of what any private body has certified. A standard is a commercial signal, not a compliance discharge.
What is the difference between a trust standard and a compliance audit?
A compliance audit examines whether you meet a specific legal requirement, usually against a defined control set, and is often performed by an accredited body. A trust standard makes a narrower, checkable public statement — commonly about identity and process — and is not accredited. Conflating the two is the most common and the most damaging overclaim in this market.
How do I choose between AI trust standards?
Apply the six tests above to each one, and weight the register test highest. Then ask the question that decides it: if this body wanted to withdraw my status tomorrow, could it, and would anyone checking me find out? If the answer is no, the standard cannot help you either, because it cannot hurt anyone.
Is a standard with few participants worth adopting?
Sometimes, but be clear-eyed about what you are buying. A standard nobody has adopted is a proposal, and its signal value to your customers is close to zero at the start. Its value is in the mechanism and the timing — being early in a register that later matters. Anyone who tells you a young register is already a market signal is selling.