Meridian Line's safety-review summarizer occasionally echoes a crew member's home address and passport number straight out of an incident narrative into its summary, because the source text contains that PII verbatim. Where should the fix be applied so the summarizer stops surfacing that PII in its responses?

The correct fix is an output guardrail: a check that inspects the model's generated response before it reaches the user and redacts or blocks recognizable PII patterns (names tied to sensitive context, passport numbers, addresses) regardless of why the model produced them. This works because it is a deterministic, auditable control point that does not depend on the model 'choosing' to behave differently. Retraining the foundation model on a scrubbed corpus is the closest-looking trap — it might reduce the problem over time but is slow, expensive, and does not guarantee the model won't still surface PII it saw elsewhere in a retrieved chunk at inference time, since the retrieved narrative itself still contains the PII. Lowering temperature changes how much the model paraphrases versus quotes, but it is not a reliable safety control — a low-temperature model can still copy sensitive spans verbatim, and a guardrail is needed regardless of temperature setting. Switching to an embedding model is a non-sequitur: an embedding model produces vectors for retrieval, it does not generate summaries, so it cannot perform the summarizer's job at all. On the exam, remember that guardrails are applied at the input boundary (before the model sees a prompt) and the output boundary (before a response is returned), and PII leakage from retrieved content is an output-guardrail problem even when the root data quality issue also needs separate remediation.