Option A is correct because the safest guardrail architecture is defense in depth, not dependence on one control. Anthropic’s guardrail guidance explicitly recommends combining multiple safeguards: input screening and validation, hardened system instructions, safe handling of untrusted tool content, least-privilege permissions, output screening, monitoring, and red-team testing. The important engineering property is independence: if one layer misses an attack or unsafe request, another layer can still prevent harmful behavior or block delivery.

Option B places control only at the output boundary. Human review can be useful for high-risk cases, but it does not protect tool execution, data access, prompt injection, or other failures that can occur before final output. Option C relies on a single system-prompt control; system instructions are probabilistic and cannot provide complete enforcement against adversarial or malformed inputs. Option D is weaker still because model-level safety alone does not enforce application-specific policies.

Therefore, A best matches Claude Developer security guidance: layer preventive, detective, and enforcement controls so there is no single guardrail whose failure exposes the application. Relevant Study Guide topics: guardrails, prompt injection, input validation, output screening, least privilege, defense in depth, and production monitoring.

===============