Work / AI systems
Cypher
Prospect research by an AI system that is only allowed to look inside its own tenant, and has to ask a person before anything leaves.
Prototype · internalprivate sourcesynthetic tenant data only
The system that must ask permission
- Tenant checkEvery retrieval needs a validated tenant. Without one, it fails closed.
- RetrievalEvidence comes only from that tenant; crossing tenants needs an explicit admin mode.
- GuardrailsPrompt-injection and unsupported-claim checks are pinned as tests.
- PolicyPolicy and SHIELD runtimes decide what may be proposed.
- A personAn operator approves. Live sending is gated and not enabled.
Measured, not asserted
The phase-3 retrieval benchmark, reviewed by a person and audited claim by claim.
30
Hazina cases
Tenant A
Benchmark30
Jamii cases
Tenant B
Benchmark20
cross-tenant cases, where leaking is the failure
Boundary
BenchmarkWhat this page does not show
- Production use: nothing is operated, and no message is sent to anyone.
- Real customer data: tenants are synthetic.
- A single score: the benchmark is reviewed case by case, not reduced to one number here.