Skip to content
AI Risk ResearchIndependent monitoring and live experiments
Contact meSupport this project
Built for Meta's Global AI Developer Hackathon

Safety method

How handoffs are tested

Routing quality is evaluated separately from answer quality, with explicit failure labels and no invented evidence.
01

Leak tests

Canary strings and field allowlists detect whether private context crosses the approved boundary.

02

Hallucination tests

Claims are checked against task fixtures and cited evidence before a run can pass.

03

Routing tests

The evaluator checks whether local work stayed local and whether cloud escalation was necessary, minimal and correctly returned.