Atomic & composite policy enforcement tests
Each scenario is a self-contained test of one policy axis. Click any card to view the full fixture, expected outcome, and run it against the live deployment.
Single-policy isolation tests. Each scenario exercises one rule in isolation.
Approved request, evidence recorded
Happy path: valid JWT, allowed everything
A permitted AI request passes every policy check and executes — and the decision is still recorded and signed. Compliance isn't only about blocking: every allowed action carries proof it was allowed.
App exceeds its permissions
JWT missing the route's required scope
An application asks the AI to do something it was never authorized for. Without runtime checks, over-permissioned apps silently expand what AI does with company data.
Data leaves the approved region
JWT home_region does not match edge region
A request tries to route to an AI model outside the approved geography. For EU personal data this is a cross-border transfer risk — the classic data-residency violation.
Unknown business unit uses AI
Tenant not configured
A tenant with no configured policy tries to use AI. Denying by default prevents 'shadow AI' — usage nobody set rules for.
Unapproved AI model
Model not in tenant allowlist
A request targets a model the organization hasn't approved. Model allowlists are how companies control which providers see their data and which model versions make decisions.
Stale credentials used
Expired JWT
A request arrives with expired credentials. Expired-credential acceptance is how leaked old tokens become breaches; here it's refused before the AI ever runs.
Credential used in the wrong place
Wrong JWT audience
Valid credentials for one system are replayed against another. Catching audience mismatch stops cross-service token reuse.
Credentials from an untrusted source
Wrong JWT issuer
The request's identity was issued by an authority the organization doesn't trust. Only pre-agreed identity providers are accepted.
Tampered credentials
Invalid JWT signature
The credential's signature doesn't verify — it was forged or altered. The request is rejected before reaching any AI system.
Anonymous AI usage blocked
Missing Authorization header
A request with no identity at all. Every AI action must be attributable to a known identity — anonymous usage means an audit trail with holes.
Unmapped endpoint refused
Unknown route
A request to an endpoint no policy covers. Deny-by-default means new paths can't silently bypass governance.
No composite scenarios defined yet.