The charter-desk clause assistant occasionally answers a clause-number question with a number that sounds plausible but does not exist in the retrieved charterparty text. Which guardrail design most directly prevents this from reaching a broker?
Select an answer to reveal the explanation.
Short Explanation
You can't reduce a fabricated citation with temperature or length limits, those don't check the fact against the source. The only thing that actually catches an invented clause number is comparing what the model said against what was actually retrieved, and refusing to pass through anything that doesn't match.
Full Explanation
Fabricated citations are a grounding failure: the model produced a specific factual claim, a clause number, that is not supported by the retrieved context. The guardrail that directly addresses this checks every cited clause number in the draft response against the retrieved passages and blocks or rewrites the response when a citation cannot be verified there, which stops the ungrounded claim before it reaches the broker. Lowering temperature to zero makes generation more deterministic but does not make it more grounded, a model can confidently and consistently produce the same wrong clause number at temperature zero just as easily as at a higher temperature. Limiting response length reduces how much text is generated but does not target the specific failure of an unsupported citation, and a short response can still contain a fabricated number. Requiring a user acknowledgment shifts responsibility onto the broker without actually reducing the rate of fabricated citations reaching them, it is a disclaimer, not a grounding check, and does not stop the underlying problem. Exam framing: mitigating a specific class of unsafe or ungrounded output calls for a guardrail that inspects the actual claim against actual source evidence, not a generic decoding or UX-level control.