A publisher's editorial team built a Q&A assistant grounded on their style guide and now wants an automated way, before wider rollout, to check whether the assistant's answers actually stay consistent with the retrieved style guide passages rather than drifting into unsupported claims. Which Microsoft Foundry capability addresses this directly?
Select an answer to reveal the explanation.
Short Explanation
What the team wants to catch is drift, answers that sound fine but wander away from what the style guide actually says, and there is a specific measurement built for that: a groundedness score that checks how closely a response sticks to the passages it was supposed to be pulling from. A table comparing what different models cost to run is a budgeting tool, not an accuracy check. Screening for offensive or violent language catches a completely different kind of problem; a response can be perfectly polite and still be making things up. And simply letting responses run longer does not make them more faithful to the source material, if anything it gives more room for unsupported claims to creep in. A metric that directly compares the answer against the retrieved source text is what actually answers the team's question.
Full Explanation
The correct answer is D. Microsoft Foundry's built-in evaluation capabilities include metrics such as a groundedness score, which specifically measures how consistent a generated response is with the retrieved source passages it was supposed to be based on, directly addressing the editorial team's need to catch answers that drift from the style guide before rollout. Option A is incorrect because the model catalog's cost comparison helps with budget decisions when selecting a model, but it says nothing about whether a given assistant's answers are factually consistent with retrieved content. Option B is incorrect because Content Safety's harm categories screen for offensive or dangerous language, which is a different concern from whether an answer is unsupported or invented relative to source material; an answer can be perfectly polite and still be ungrounded. Option C is incorrect because allowing longer responses does not check accuracy or consistency with source content at all, and could even give the model more room to introduce unsupported claims. The groundedness metric within Foundry's evaluation tools is purpose-built for exactly the consistency check the editorial team needs before a wider rollout.