Which guardrails belong on the output path of a customer-facing endpoint?

An **output guardrail** runs after the model generates text but before that text reaches the user. Its job is to inspect or clean what the model itself produced. Three of the options do exactly that. Content moderation classifies the generated text for unsafe content. PII detection scans for leaked personal data — names, emails, card numbers — and redacts it before the response goes out. Policy and format checks enforce your rules: no competitor mentions, no banned claims, and a valid schema if the response is structured. The other two aren't guardrails at all. Raising the model's **temperature** is a generation setting that adds randomness; it changes how the model writes, but it inspects nothing and actually increases risk. 'Caching the GPU's power draw' is nonsense — it's not a safety control of any kind. The key idea: input filtering alone can't catch these problems, because the harmful content originates in the model's output. You need a check on the way out.

Output-path guardrails sit between generation and the user, inspecting or sanitizing what the model emits. The three correct options all fit that definition. **Content moderation** runs a safety classifier over the generated text. PII detection and redaction scans the output for personal data and masks it before the response leaves. Policy and format checks enforce business rules and structural validity. Examples are no competitor mentions, no disallowed claims, and a valid schema. The reason these belong specifically on the *output* path is that the harm originates in the model's own generation. Input filtering catches bad prompts, but it can't catch a model that hallucinates a fake policy, leaks PII it inferred, or drifts off-format. Only a check after generation sees what was actually produced. The two distractors fail the definition. Raising **temperature** is a generation parameter; it changes the sampling distribution to add variety but inspects nothing. More randomness tends to increase unsafe or off-format output, not reduce it. Caching the GPU's power draw isn't a control at all — it's a nonsense option mixing an unrelated metric with an unrelated mechanism. A guardrail must observe content and gate or transform it; neither distractor does either, so the answer is the three genuine inspection and sanitization steps.