xChk TVP | Trust ◦ Verify ◦ Proof
Explained

An emergency stop for AI agents

One call halts a plate, a host, an organisation or a whole registry. The halt is signed, verified on the host before any permission is read, and measured.

Every serious machine has a stop that works regardless of what the machine is doing. Agents mostly do not. When one misbehaves, the options are to kill a process you may not own, revoke a key the agent may have cached, or ask the vendor. None of those is a stop; they are hopes. An emergency stop for agents has to halt the thing itself, quickly, at a scope you choose, and leave proof that it did.

How the xChk stop works

A halt is a registry-signed order at one of four scopes: a single plate, every plate on a host, every plate in an organisation, or every plate the registry has issued. The gate on each host verifies the signature and checks for a halt before it reads any grant. A halted agent therefore does not get to the question of whether an action is allowed; the answer is already no, and the refusal is recorded. Lifting the halt is the same kind of signed order, so the record shows who stopped what, when, and who started it again.

Why it is verified on the host

A stop that depends on the registry being reachable fails exactly when you need it most. The order is signed so the host can check it without a live call, and the gate fails closed when it cannot reach the registry at all. Either way the agent does not run unguarded. This is the same principle as the receipts: the host trusts a signature it can verify, not a server it has to believe.

Measured

A stop is a claim until it is timed. On the November 2026 release the halt benchmark sent 20 halts and 20 of 20 took effect, with a median of 9 milliseconds and a 95th percentile of 221 milliseconds between the order and the first refused action. The method and the results sit with the rest of the benchmarks. The stop also appears in the AGSC self-assessment, where kill-switch and revocation controls are rated against what the system actually does.

What the dashcam adds

A stop tells you the agent halted. It does not tell you what happened before. The dashcam replays a plate's sealed record in order and names the actions it cannot account for, so the person who pulled the stop can see what they stopped, and an outsider can check that the account is complete.

Scope is the design decision

The four scopes exist because incidents come in sizes. A single agent gone wrong wants a plate halt; a compromised laptop wants a host halt; a vendor incident may want everything in an organisation stopped at once; and a registry operator facing a key compromise needs to be able to stop every plate it ever issued. Each scope is one call, and each is reversible by the same path.

Where it fits with the rest

Patents pending. Numbers above are from the published benchmark runs; they are re-run on every release and the page is updated when they move.