xChk TVP | Trust ◦ Verify ◦ Proof
Explained

Authorization receipts for AI agents

A record of who allowed what, sealed at the moment it happened, that anyone can verify afterwards without trusting the vendor or the deployer.

When an AI agent sends an email, pays an invoice or runs a command on a server, two questions follow it: was that allowed, and who allowed it? A log answers neither. A log is written by the same software that acted, it can be edited by whoever runs that software, and it says nothing about whether a person was involved. An authorization receipt is the answer to both questions in a form a stranger can check.

What a receipt is

In xChk every gated action produces one record. The record names the agent's plate (its licence), the tool and area the action fell into, the exact command or request, the decision (allowed by a grant, held for a person, refused), and when a person answered, who they were and how they proved it, usually a passkey on their phone. The registry signs the record and chains it to the one before it by hash, so a record cannot be altered or removed without breaking every record after it. Batches of records are anchored to public timestamping services, so the chain's existence at a given time does not rest on the registry's word either.

The receipt is the public face of that record: a page per approval that shows the action, the decision and the approver, and a signed statement that can be checked offline. The statement format and its checker are open; the checker needs no shared secret and makes no call to the registry.

What it is not

It is not a transcript of the agent's reasoning, and it is not a claim that the action was wise. It is evidence that a specific person authorized a specific act at a specific time, or that no person did and the system said so. That distinction matters when something goes wrong. An insurer, an auditor or a counterparty does not need to understand the model; they need to know whether the controls were on and who answered.

Why offline verification matters

A verifier that trusts the server proves nothing. If the only way to check a receipt is to ask the registry that issued it, the registry can say whatever it likes. So the receipt carries enough to be recomputed: the record hashes, the seal, the chain position and the anchor. The Black Box report on this site does exactly that in your browser, recomputing two record hashes and one seal and confirming the chain is intact, and it also shows a red-team drill on the same plate where the recorder was deliberately withheld. A missing record is itself evidence, and it is shown rather than hidden.

Two registries, one format

The receipts on this site come from two independent registries with separate trust roots, in.xchk.io and xchk.me. A receipt from either verifies the same way. That is the point of the interchange paper: accountability has to survive crossing from one operator to another, the way a driving licence is readable in a state that did not issue it.

Where receipts fit with the rest

Measured, not claimed

Receipts are only worth what the gate that produces them is worth. The benchmark page shows the control-plane-integrity results, including the tamper and anchor checks, and the external suites the gate has been run against. The AGSC self-assessment maps receipts to the controls in that catalog.

Patents pending. The format and the verifiers are open so that a receipt means the same thing whoever issued it.