Recorded verification · November 6, 2026
What the deployed system did
These results come from a dated release trial using real Auth0 identities and AgentGuard’s simulated ledger. The public walkthrough uses separate scripted examples.
22 live assertions passed in one release trial
These are checks within one trial, not 22 independent runs or a load test.
- Human review and execution
- $755.00 simulated refundOwner approved after passkey reauthentication; exactly one simulated receipt and one human-review audit event.
- Approval replay
- Second execution attemptDenied. The same approval still has only one receipt.
- Request retries
- Identical and changed payloadsIdentical retry returned the original result; a changed amount with the same request ID was denied.
- Cancellation
- Pending request cancelledOne cancellation audit, zero receipts; execution denied.
- Pause and resume
- Simulated one-cent refundBlocked while paused with no partial request; a new request succeeded after resume.
- Permissions
- Real Auth0 machine tokensSales could read CRM, could not issue refunds; the machine administrator could not approve.
Separate local evaluation
Same requests, three sets of controls
Each scenario checks the resulting receipt count and total amount. AgentGuard ran its actual SQL migrations against an empty in-memory database; both comparison baselines are intentionally limited local models.
- Unguarded
- 1 of 10
- matched the expected effect
- Scope and per-request threshold
- 3 of 10
- matched the expected effect
- AgentGuard database controls
- 10 of 10
- matched the expected effect
7 SQL assertion files also passed, including expiry and rollback checks. This single-connection run uses synthetic identities and does not test live authentication or concurrency.
Revisions tested
- Frontend commit
- 5c469b4234451fb192c8692113bb0e8ceaf26f74
- Backend commit
- 65c3633fdbf1e3b43df9b17eb5a4946ef03a409b
- Database
- Migrations 001–008
Reproduce the checks
Start with the backend setup, refund contract, and database tests. The release report separates local tests, database assertions, and the live trial.
What this evidence covers
The owner performed approval. Cancellation and pause/resume used the authenticated administrator interface during verification. No real payments moved. The application trusts its privileged server and database operators, so the audit trail is not a cryptographic guarantee against those operators.
Separate local MCP tests exercise the real verifier with signed wrong-issuer, wrong-audience and expired tokens. They use fixture discovery and public keys. The full negative-authentication browser matrix and a fresh hosted Supabase/Auth0 rehearsal remain open.
Read the hosted rehearsal guide
Return to the walkthrough