Trust
We try to break it on every release.
Our own testing on a development stack. A release that fails these doesn't ship.
- 36 of 36end-to-end scenariosRelease of 1 Nov 2026
- 11 of 11prompt-injection attacks in scope, heldRelease of 1 Nov 2026
- 14 of 14race, failure and load casesRelease of 1 Nov 2026
- 12 of 12checks that one real supplier bank change ran once, matched and was signedRun separately, 27 Oct 2026, in a payment provider's test mode
Check a receipt yourself
Signature (Ed25519) Ns_M28aaIN3EonbO…jraRhJAA
Sample receipt, signed with a throwaway demo key. Real receipts carry more fields, and our separate verifier also checks the chain of records behind them.
Press Verify and your browser checks the signature itself; nothing is sent anywhere. Then change any value, even one digit, and check again.
Not checked yet.
| What we test | What must hold |
|---|---|
| A real action, in test mode | Runs once, matches at the target, is signed, and can't be replayed. |
| Prompt-injection attacks | Even when the model is fooled, nothing runs without a person. Delegrity is tested as if every attack had worked. |
| Races | One request sent 12 times makes one action and one decision. Two approvals or two executors make one effect. |
| Failures | With a service or the target down, nothing reports false success and nothing runs twice. |
| Revocation | Authority taken away stops execution, even with valid credentials and a grant in hand. |
| Tenant isolation | One tenant can't read or change another's data, including under load. |
| Tampering | An altered receipt or evidence record fails independent verification. |
| Different AI models | A weak model is reported as a model error. It never becomes a Delegrity pass. |
No production customer evidence or independent certification yet.