What's this about?
The benchmark tests routing, fares, disruptions, accessibility and adversarial requests in six real metro systems, requiring structured tool calls and machine-renderable outcomes.
What exactly is a policy layer?
It is the component that decides what action an application should take from the current request and rules; here, the model is the decision-maker behind a kiosk interface.
Why is it worth attention?
A small specialized model can beat a much larger general model on a narrow deterministic score and run in 2.6 GB. Six semantic components use an LLM judge, and the lead changes on the combined ranking.