Skip to main content
A single Proofable job records five things in order. They are who the human is, which agent ran, what that agent was allowed to do, the decision to dispatch it, and what it actually returned. Each is signed and carries a qHash, and the signed decision commits to the delegation and agent-identity that produced it, so the chain can be rebuilt from the records alone on any machine.

The chain

The joins are hashes

A database row can point at another row only inside that database. These records point at each other with hashes, so the pointer survives export.
  • delegationQHash and agentIdentityQHash inside the signed decision name the authority and the agent that produced it.
  • argsDigest is a sha256 over the request as sent, so a reader holding the decision can recompute what was authorized without trusting a log.
  • The job record carries the terminalQHash of the signed receipt for that execution, and the same delegationQHash and agentIdentityQHash the decision named.
  • statusReadAtMs and maxStatusAgeMs inside the signed decision say when the grant was read and how old that read may get while the job runs (60 seconds), so the age is checked from the record.
  • The signed receipt carries authorityDecisionQHash, the decision the job ran under. A reader goes from the receipt to the decision with no other source.
The decision is recorded before the job exists, so its own jobId is null. The receipt points back to it instead. An external reviewer or another agent can audit a run they did not make without reading our database.

Run it end to end

The reference kit drives all five steps against the hosted service and writes the records to a folder on your machine. It needs your own access key, because it runs real agent jobs against your plan.
verify prints:
node aad.mjs controls runs/<run-id> then alters a copy eight ways (a case dropped, duplicated or flipped, a citation removed or stripped, genuine records swapped between cases) and passes only if verify rejects every one and accepts the unaltered copy. Both passing is the acceptance criterion for a run. verify needs no network and no account. The eight cases cover an allowed job, an action outside the delegation, an approval that arrives, an approval that never arrives, revocation before dispatch, expiry before dispatch, reuse of a revoked grant against a valid control, and revocation while a job is running. Their expected outcomes are listed in the kit’s README. To run the same cases in your own CI, use the GitHub Action from the same kit. It fails the build when a record does not verify.

Check a record without taking our word for it

It recomputes the qHash from the envelope and recovers the signer from the signature, with no network call and no account. The CAIP-380 profile documents the canonical form, so you can write the same check in another language, and the signing format covers wallets that sign with something other than an EOA. Published runs, with the records and a script that verifies every envelope in them, are on the Benchmark page.

What the chain does not cover

  • The effect at a target you control. Records state effects as the service observed them at the job boundary. A readback taken at a target outside that boundary is a separate observation, and the kit’s URL fetch is where to add one.
  • How long a revocation takes to reach the check. A job running when its permission is revoked is stopped with a typed reason, and the next dispatch is refused. No bound on that delay is claimed.

Build on it

  • MCP setup for an agent that calls the service directly.
  • JavaScript SDK for the same calls in your own code.
  • Sell access to require a proof before your product serves a request.
Last modified on November 17, 2026