This question addresses important concerns in the field of AI ethics and compliance, particularly as...

The rejected options are weaker because keyword filters and one-time prompt disclaimers do not enforce policy under prompt injection, ambiguous requests, or regulated-domain escalation paths. RAGAS-style metrics can support guardrail evaluation but cannot independently cover every safety issue. It should be one measurement layer, not a total compliance solution. Option A is the correct engineering choice because the requirement is not just “make the model answer,” but control the execution surface. The selected option specifically A states “RAGAS cannot evaluate all safety aspects independently but provides metrics like Topic Adherence and Agent Goal Accuracy that serve as guardrails.”, which matches the operational requirement rather than a superficial wording match. In NVIDIA terms, Guardrails are most effective when paired with evaluation, red-team prompts, and audit metadata so coverage gaps become visible. The durable control mechanism is guardrail coverage that is tested against observed failures and adversarial prompts rather than assumed from policy text. For certification purposes, read the question as asking for controlled autonomy, not raw LLM creativity.