Seal

Seal · for reviewers

Three things a signed record can be asked to prove. One of them is not solved.

Everything below runs on a real production receipt, signed by Aqta's gateway on 13 August 2026. Break each one yourself. The point is not that the first probe passes. It is where the third one stops.

01 Authenticity

Solved

Do exactly these fields verify against the published key, and has anything changed since they were signed? Edit any value below. The signature covers the canonical bytes of all eleven fields, so changing one character changes what was signed.

Solved means this and nothing wider: nobody has altered the record since it was signed. Probes 02 and 03 are what that leaves untouched.

attestation_id
520a3601-0d94-4b32-9619-a512ca853a94
cost_prevented_eur
0
model
claude-sonnet-4-6
org_id
ab451951-318a-4527-be74-7420f9c4356c
outcome
ALLOWED
policy_applied
budget_guard, loop_guard
public_key
9Y3Eiq6V8QjRDUM5nPqSwKIOPQaoEU4SbagfYFdvWa4
request_hash
577f7a94d8985770f5e559de72b75dfe3f7a54c00dc07aa299946ac381586cf9
timestamp
2026-08-13T01:27:11.840747+00:00
trace_id
c2a3e22f-5753-4291-881a-533a86b14679
v
1
signature
65rPtNi4jeNe7D4ahoS5FpemS2Z54C3VcLVJ7lIWWUwj9b4muw4ZsHEmgQf_fiXT9EDANddpM_eXCITFeXmWBg

Click any value to edit it.

bytes matchUnchanged since signing. This is the receipt the gateway produced.

What this page does and does not do. It shows you which bytes the signature covers, so an edit is visibly outside them. It does not run Ed25519: to check the signature itself, run npx aqta-verify-receipt against the published key, offline, with no account.

02 Correctness

Partial

The record says a particular model ran under particular rules. Does the signature establish that this is what actually happened inside the machine?

FieldWhat it establishes
modelThe gateway stated claude-sonnet-4-6
policy_appliedThe gateway stated budget_guard, loop_guard
request_hashA SHA-256 the gateway computed over the request
signatureThe key holder asserted all of the above

Press the button to see what the signature still says.

A signature is an assertion by a key holder, not evidence of execution fidelity. It proves who said it and that nobody edited it afterwards. Binding the record to the computation that actually ran needs hardware attestation; ours is laboratory work at TRL 3, outside production, as our own threat model states.

03 Completeness

Open

Here are the decisions one lender sealed on 13 August. Every one verifies. One consequential decision that day was never written down at all. Find it.

TimeAttestationOutcomeSignature
01:27:11520a3601ALLOWEDvalid
04:02:56c81be40fALLOWEDvalid
09:41:087d2aa914BLOCKEDvalid
11:19:33e0c47b52ALLOWEDvalid
15:52:471f9d3ac8SUPPRESSEDvalid
18:07:209ab6f701ALLOWEDvalid

Every row above checks out. That is the problem.

Nothing we know of closes this. An append-only log over issuance narrows the window in which a record can be withheld, and narrowing is not closing. We publish this as open, not solved and not claimed.

Receipt 520a3601-0d94-4b32-9619-a512ca853a94, signed 13 Aug 2026 under key 9Y3Eiq…. Verify it yourself: the record formats, both verifiers and every published conformance vector are public, and a reviewer countersigns one at review.aqta.ai. Signing is production. The review flow is a prototype and says so. Back to security.

Now taking pilots.

A free 30-minute evidence review first, then up to five working days on one workflow: one review path, one evidence pack, judged against your reviewer’s own criteria.