Free Claude Certified Architect Professional (CCAR-P) Sample Questions

D. Embeddings capture meaning but handle rare exact tokens like error codes poorly. That is why the correct article never reaches the candidate set for code-bearing queries. BM25 matches the literal token ERR-4471 inside the prose, and fusing its ranked list with the embedding results keeps the semantic recall that already works for the conceptual part of mixed queries. The deciding details are that the codes live only in free text and that queries combine a code with a natural-language question. Why the others are wrong: An exact-match lookup would win if the codes were stored in a structured field and queries were pure identifiers, but here no such field exists and the queries also carry conceptual intent. Contextual chunk prefixes help when chunks lose meaning once separated from their source document, but these articles are already short and self-contained, and the failure is exact-token matching, not missing context. A reranker helps when the right article is in the candidate pool but ranked too low, yet here it is missing from the top 150 candidates, so there is nothing for the reranker to promote. Source: CCAR-P exam guide, Domain 3.

D. Model selection should use the smallest model that meets the quality bar. Larger tiers only earn their extra latency and cost when evaluations show a capability gap. Here the pilot showed the fastest tier matching larger tiers on every task, including the hardest rule questions. Sellers also need near-instant responses at very high volume, so the fastest, lowest-cost tier is the right choice. Why the others are wrong: The mid-tier model everywhere would be the right default if the pilot had shown the fastest tier falling short, but here it adds latency and cost with no measured quality gain. A mid-tier default with escalation suits workloads with a measurable tail of hard cases, but the evaluation found no cases where larger tiers outperformed the fastest tier. Sending listing-rule questions to the top tier would make sense if those questions needed deeper reasoning, but the pilot showed the fastest tier already handles even the trickiest ones at the quality bar. Source: Claude docs: Choosing the right model.

C. The deciding detail is that the key is already exposed in a public commit, so anyone who copied it can call every tool, including the safety-data assistant, until the key stops working. Revoking the key is the only step that immediately cuts off that access. Issuing new keys per tool then restores service and can be combined with tighter scoping afterward. Why the others are wrong: Workspace separation with scoped keys fixes the underlying least-privilege gap and would be the best answer if no leak had occurred, but the leaked key keeps working until it is revoked. A secrets vault is the right fix for keys stored in plain-text config files, but the key is already public, so moving it now does not stop anyone who has copied it. Reviewing usage data is the right step when investigating possible misuse, but it only reports on the exposure while the leaked key remains live. Source: CCAR-P exam guide, Domain 5.