An architect must design a Claude-based system that, when given a user question, must never answer from memory alone — it must always cite a specific retrieved source document for every factual claim. Default Claude behavior allows the model to draw on training knowledge even when retrieval is available. What prompt engineering approach most reliably enforces the citation-mandatory policy?

B is correct because enforcing citation-mandatory behavior requires both a structural constraint (citation block schema that makes unsupported claims syntactically detectable) and an explicit fallback instruction ('state No source available rather than relying on training knowledge'). This two-part approach prevents both uncited claims and silent use of training knowledge. Post-generation validation of citation block coverage provides a quality gate. A is wrong because the instruction 'Always cite your sources' is too vague and does not create a detectable schema that can be validated programmatically — Claude may partially comply or interpret 'sources' loosely. C is wrong because temperature controls output sampling distribution and has no effect on whether Claude draws on training knowledge versus retrieved context — the same prompt at temperature 0 will still produce training-knowledge-derived claims if the retrieval context is incomplete. D is wrong because fine-tuning is not available as a self-service option for Claude, and even fine-tuned models can violate behavioral constraints probabilistically without the structural schema enforcement described in B.