Accountable AI for consequential decisions.
Who cleared this?
Controls for AI that hold when challenged.
every line here is checkable
- Signing in production, with a published key
- Open verifiers on npm and PyPI
- Record format filed as an IETF Internet-Draft, individual submission
- Public log of signed records
What Seal checks
An action nobody cleared can look perfectly reasonable.
Detection asks what looks dangerous. Seal asks what a person cleared.
- A supplier page says the bank details have changed.
- A note asks for one payment to be split in two.
- Yesterday’s approval, presented again.
- A release that is not the one signed off.
one payment
Cleared: €4,800 to Harbour Print, the account on the invoice.
Proposed: the same €4,800, to the new account on the supplier’s page.
RefusedThe approval named one account, not any account.
and the refusal is signed
Each can arrive with a good explanation. Seal does not judge how a request sounds. Before the action runs, it checks the exact amount, payee or version against what a person cleared, and signs the outcome either way.
Pre-registered evaluation
Can an AI agent act beyond the authority it was given?
Authority Arena tests AI agents under pressure, comparing different controls against unauthorised actions and measuring whether legitimate work still gets done.
Try it
AI agents can now reason and act. They shouldn't be able to give themselves permission.
Propose an action. Seal checks it against a demo rule set and signs the decision.
A record shows the decision Seal signed for the proposed action, unchanged since. It does not show that the action ran, or that every action went through Seal. Seal enforces at the gateway you route actions through; an action that bypasses the gateway is neither checked nor signed.
an agent proposes
agent release-agent
tool git.push_production
args {"repo":"payments-api","ref":"main"}the signed record lands here
Use cases
Logs tell you what your system recorded. Seal gives you a record an outside reviewer can verify.
AI vendors selling into regulated buyers
“How do we verify what your AI decided in production, and under which policy?”
Financial services
“Why was this application declined on the fourteenth of March?”
Insurance
“Did each of these decisions clear under the policy that was in force when it ran?”
Healthcare
“What exactly did it decide, and can we show it?”
Checked rather than believed.
Pilot programme for teams deploying AI in consequential workflows. Two weeks, fixed scope: one workflow, one review path, one evidence pack.