A generative answer is relevant but unsupported by the configured knowledge. Which evaluation should be added?

1. Only response length
2. Only channel count
3. Grounding and task correctness against authorized evidence
4. Only a successful HTTP response