A database-grounded assistant gives an answer that conflicts with the retrieved records. Which TWO checks should guide investigation? Select TWO.

1. Whether the answer contains more words than the source rows
2. Whether generation faithfully uses the authorized evidence and task instructions
3. Whether the model can bypass permissions to find a preferred answer
4. Whether the retrieved evidence actually answers the user's question