For teams accountable for what AI is allowed to do.

Controls for AIthat hold when challenged.

A loan, a diagnosis, a claim. Cleared before it runs, signed when it happens, and checkable offline.

AI acts. Keep the receipt.

Why a seal

For three thousand years, when someone had to trust a thing they had not witnessed, we pressed a seal into it.

  • Who authorised this?
  • Has it changed since?
  • Can a stranger check it?

AI has learned to act. It deploys code, moves money, and writes to records that matter. The questions have not changed. The object has to. When the actor is an agent choosing its own tools, that gets its own record: ACTION-v1, decided before the tool runs.

Checking in your browser

Sample record, signed with the published demo key. A real production refusal is in the refusal ledger. Two independently written checkers, one TypeScript and one Python, agree on every published conformance vector.

  "outcome": "BLOCKED",
  "model": "gpt-4o",
  "signature": "uZCP7STzFgHXbyIK1oKoZ5Zn8-…"

Authority can be delegated. Trust should remain checkable. Seal sits alongside the AI systems you already use and gives the teams running them a record they can carry into security review, audit and customer assurance.

Check a full record

Seal by Aqta

The next generation of "who cleared this?"

  1. Clear
  2. Seal
  3. Verify

Clear. Seal. Verify.

A loan, a diagnosis, a claim. Seal records what the AI was cleared to do before it runs, signs what happened, and the record checks offline against a published key, years later, on the day someone asks.

Try it: app.aqta.ai/verify

SealSample receipt
When
Before execution
Policy
Personal data check
Outcome
Allowed

Signing

Writing the signed record

Signed record

When the answer matters

Months later, someone asks what happened.

When a decision matters to someone outside your system, they need something they can check for themselves.

With evidence that holds, consequential AI clears security review, procurement and audit, and ships. See a sealed record

No regulation asks for cryptographic receipts. What a reviewer asks is whether the answer holds when they check it themselves.

DORA, 2025Moffatt v Air CanadaFINMA 08/2024Agent intrusion, July 2026The full picture →

  • One pack, one link
  • Nothing to install
  • Still verifies years later
  • The verdict is signed too
docket · the record a review ends with

“Did this decision clear under the policy in force when it ran?”

evidence
receipt sealed · pack exported
signature
verified in the reviewer’s browser
reviewer
no account · published key only
clock
question → sealed answer
ACCEPTED · on the record

Our thesis

Once AI has authority over money, patients and code, proving what authority it was given becomes infrastructure.

AI providers make increasingly capable systems possible. Seal sits alongside them, so a consequential decision carries proof of what it was cleared to do. Checked rather than believed.

Read the manifesto

One Seal path

Seal one workflow. Keep the receipt.

Bring a loan, a claim or a clinical decision. We clear it to ship, and the record travels with it.

Checking a record is free, always. You pay for making one at the moment the decision happens, and for standing behind it afterwards.

  • Start where a person already signs off
  • Scoped engagement
  • Reply in 4-5 business days
  • Offline-verifiable trail

Product notes by email