Prove your AI agent was allowed to do that.

AI Judge decides whether an agent can spend, release funds, or complete a task — and returns a signed receipt that anyone can check. No AI in the decision itself, so the answer never drifts.

See a real decision Read the API docs

Guardrails aren't evidence

Most controls for AI agents are another model checking the first one. Ask twice, get two answers — which is worthless the moment there's a dispute, a chargeback, or an auditor. Enforcement for money and regulated data has to sit outside the model entirely, and it has to leave proof behind.

A real decision, verified live

This is fetched from the live API right now — not a screenshot. The same check runs for anyone, with no account.

Loading a live decision…

Open the full receipt, edit a single character of it, and verification fails. That's the whole product.

How it works

  1. Send the evidenceThe agent's identity, its spending limit, its reputation — whatever your rules depend on.
  2. Your policy decidesPlain rules you version and publish. Not a prompt. It cannot hallucinate a verdict.
  3. Get a signed decisionApprove, deny, or send to a human — with the reasons, a signature, and a link anyone can verify.

What you get that logs can't give you

Deterministic

Identical inputs always produce an identical verdict. Provably, not probably.

Signed

Every decision carries an Ed25519 signature. Alter one character and it fails.

Replayable

Re-run any past decision from its original inputs and prove it still holds.

Tamper-evident

An append-only, hash-chained audit log. Deleting an entry breaks the chain.

Check any decision yourself

Public verification needs no key and reveals nothing private — no evidence, no scores, no customer data. Just whether the decision is genuine.