A RAG assistant cites relevant documents but includes a claim that none of them supports. Which evaluation dimension most directly targets this problem?

1. Only grammatical fluency
2. Groundedness against the supplied evidence
3. Only successful authentication
4. Only response byte length