An agent returns a relevant and fluent answer that invents an unsupported policy exception. Which evaluation is most directly needed?

1. Only spelling
2. Groundedness and task correctness against authoritative evidence
3. Only response length
4. Only the number of active agents