The Empty Handshake

Article 50 of the EU AI Act has applied since 2 August 2026: a person must be told when they are dealing with an AI. Machine-readable disclosure — an agent card, an llms.txt, a statement on the page — is how that obligation becomes checkable by anyone rather than assertable by the deployer.

A single count cannot answer whether it is appearing or surviving. A series can, and the first year after the deadline is happening once. The predictions were written down and sealed to a public clock before the data was collected, so a reader can establish that they predate it without taking Vigilia's word for anything.

Not one of the reachable EU AI Pact signatories publishes an agent card at the address other software looks for. Hundreds of hosts listed in public agent registries do. The same week, the same method, the same five addresses: the machine-readable layer exists, at scale, and almost none of it sits on the public face of the organisations deploying AI.

This corrects a wider claim Vigilia published on 20 September, when it had measured only corporate front doors and a single forum. The narrow prediction stands; the sentence around it was too wide and has been withdrawn in the protocol.

A registry asserts that an agent is healthy. Vigilia reads the same agents from outside and puts the reading beside the claim.

Of 398 agents a registry marks healthy, 393 served a valid card when read from outside. Hosts whose robots.txt asks not to be read are excluded from both sides of this number, because that is Vigilia's restraint and not their failure — counting them as failures would have reported a larger and more flattering figure.

A gap near zero is published on the same terms a large one would be. An instrument that reports only the failures it hoped to find is not an instrument.

Each is written to resolve on a date, in a number. A miss is published as a miss, in the same table as the hits.

• Agent cards on EU AI Pact signatories stay between 0 and 2.
• llms.txt on those same hosts grows from 36% to between 45% and 60%.
• Disclosure phrases on their homepages stay within 5 points of 24%.
• Agent evidence decays an order of magnitude faster than corporate evidence.
• The two populations do not converge.
• The registry claim gap stays under 10%.

Every number on this page comes from one file, and that file's fingerprint is signed into a public transparency log each week by a party with no interest in the result.

• The data, machine-readable — /census.json
• The protocol and its amendments — /sealed/census-protocol.md
• The sealed manifest — /sealed/manifest.json

Fetch the protocol, hash it, and compare it with the entry in the manifest. Then check the manifest against its Sigstore bundle and its RFC 3161 timestamp. If a file changed after it was sealed, the hashes stop matching, and that mismatch is itself the record.

A count of surfaces is not a verdict on anyone's compliance. A homepage without a disclosure phrase is not a violation: the obligation attaches to the interaction, not the front door, and the disclosure may live inside a product this instrument never opens.

Nothing here logs in, fills a form, evades a bot challenge or opens an agent's endpoint. robots.txt is honoured, and a host that asks not to be read is recorded as not read. A card is a document, and reading a document is not using the agent.

Vigilia's own domain and its own registry listing are rows in these populations. The instrument measures its owner — and its own row is excluded from the claim gap, because grading your own claim inside a measure about everyone else is not a second source.