Meridian's charter desk says the assistant sometimes answers confidently even when no clause in the retrieved set actually addresses the question. Which prompt-design fix most directly targets this specific failure mode?
Select an answer to reveal the explanation.
Short Explanation
Give the model explicit permission - really, an instruction - to say 'the clauses don't cover that' instead of quietly filling the gap with a confident-sounding guess; without that instruction, the default behavior of most models is to try to answer no matter what.
Full Explanation
An explicit instruction telling the model to admit when none of the retrieved clauses address the question, rather than defaulting to answering anyway, directly targets the described failure mode, because the problem is a behavioral tendency (answering confidently despite irrelevant retrieved content) that only a direct instruction against that behavior corrects. Simply retrieving more clauses (three to thirty) does not fix this - if none of the relevant clauses exist in the corpus for this question, retrieving more irrelevant ones does not make the model more likely to admit that, and it may even give the model more unrelated material to (mis)construe as support for an answer. A longer context window changes how much text the model can process, not whether it is instructed to admit ignorance - a bigger window with the same permissive prompt would still produce the same overconfident answers. Removing the citation requirement makes the problem worse, not better - citations are one of the mechanisms that make ungrounded answers visible and checkable; eliminating them removes a safeguard rather than fixing the underlying overconfidence. This item tests recognizing that a described model behavior problem (over-answering) requires a targeted instruction fix, not a retrieval-volume or model-capacity change.