Claude Certified Architect Professional CCAR-P Exam Questions
Question #11 (Topic: Demo Questions)
A team proposes storing all production prompts and completions indefinitely in a general-purpose logging platform accessible to the whole engineering organization, to help with debugging. The system handles customer financial data. What is the most significant governance objection?
Correct Answer: B
Explanation:
Why this is correct. The critical insight is that observability data for an LLM system is not metadata — the prompts and completions *are* the customer data, reproduced in full. That means the logging platform is now a system of record for regulated financial information and must inherit every control that applies to the source:
restricted access on a need-to-know basis rather than organization-wide, defined retention rather than indefinite, and the ability to locate and delete a specific customer's data on request. Broad engineering access to production financial data is also an access control violation independent of retention, and the two together
would likely fail an audit.
Why C is wrong. Storage cost is real but minor, and it is an economic concern rather than a governance one. It would not block the design.
Why D is wrong. Queryability is an ergonomics issue. It affects how pleasant debugging is, not whether the arrangement is permissible.
Why A is wrong. This names a genuine benefit, and building evaluation sets from production traffic is good practice — but it must happen within the same access and retention controls. It is not an objection.
Blueprint objective: Ensure compliance with regulations; analyze observability challenges.
Question #12 (Topic: Demo Questions)
A financial services client asks how the organization will demonstrate to regulators that its Claude-based advisory tool behaved appropriately on a specific past transaction. Which architectural capability is required?
Correct Answer: A
Explanation:
Why this is correct. Regulatory defensibility requires reconstructing what the system did and why, for a specific interaction, potentially years later. That means capturing the full decision context — not just the output but the inputs, the retrieved passages that grounded it, the exact prompt and model versions in force at that moment,
the tools invoked, and any human review or override. Version capture is the detail candidates most often miss:
without it you cannot reproduce the behaviour, because the prompt has since changed. Tamper-evidence and retention aligned to the regulatory period complete the requirement.
Why B is wrong. Training data is neither available nor relevant to what happened in one interaction. Regulators ask what the system did with this customer's information on this date, which is an operational question.
Why C is wrong. A vendor assurance is not evidence about a specific transaction. No provider statement can substitute for your own records of your own system's behaviour.
Why D is wrong. Multi-region availability is a resilience property. It has no bearing on demonstrating past behaviour.
Blueprint objective: Ensure compliance with regulations; analyze observability challenges and select monitoring strategies at scale.
Question #13 (Topic: Demo Questions)
An agentic system can execute actions with real-world consequences: sending customer communications, issuing refunds, and modifying inventory. Which control design best balances autonomy with risk containment?
Correct Answer: D
Explanation:
Why this is correct. Autonomy should be granted in proportion to the cost of being wrong, and the two dimensions that determine that cost are reversibility and blast radius. An action that affects one record and can be undone in seconds does not warrant the same gate as an irreversible action affecting thousands. Tiering
preserves the efficiency case for automation on the common low-impact path while placing a human at the points where errors are expensive or permanent. The rate limits and value ceilings are the essential backstop:
they bound the damage from a systematic failure that produces many individually-approved-looking actions, which is the failure mode that has caused the most severe agentic incidents.
Why C is wrong. Universal approval eliminates the automation benefit and, at volume, degrades into rubberstamping — reviewers facing hundreds of mostly-correct proposals stop reading carefully, so you pay full cost for little real oversight.
Why B is wrong. A daily report is purely detective and arrives up to 24 hours after the fact. Refunds and customer communications cannot be recalled once issued, so the control operates entirely after the harm.
Why A is wrong. Time of day is unrelated to risk. A destructive action at 2 p.m. is exactly as destructive as one at 2 a.m.; the only thing this changes is how quickly someone might notice.
Blueprint objective: Implement guardrails and safety controls; apply human-in-the-loop validation strategies.
Question #14 (Topic: Demo Questions)
An architect must define an SLA for a Claude-based system with an external client. Which set of commitments is most defensible?
Correct Answer: C
Explanation:
Why this is correct. A defensible SLA commits only to things that are measurable, controllable, and objectively verifiable by both parties. Availability and latency percentiles are standard and monitorable. The important nuance is how quality is expressed: not as a promise of correctness, but as a target measured on an agreed evaluation set, which makes the commitment testable and gives both sides a shared definition of what "good"
means. The stated review cadence acknowledges that inputs and expectations evolve, providing a structured path to revisit the target rather than an argument later.
Why A is wrong. No probabilistic system can guarantee 100% correctness, and committing to it creates a term breached on day one. It also signals to a sophisticated client that the architect does not understand the technology.
Why D is wrong. Promising to eliminate human review forecloses a control the design may require for high-risk actions, and would force the team to remove a safety mechanism to satisfy a contract.
Why B is wrong. "Most advanced available" commits to a moving target set by a third party, may conflict with cost and latency obligations in the same agreement, and confuses an input with an outcome. Clients should be promised results, not component choices.
Blueprint objective: Manage stakeholder feedback loops and expectation alignment (including SLAs).
Question #15 (Topic: Demo Questions)
A Claude system has been live for six months. Which practice best reflects sound lifecycle management in the operate-and-iterate phase?
Correct Answer: B
Explanation:
Why this is correct. Operate-and-iterate is a disciplined cycle, not a state of rest or a state of constant churn. Reviewing monitoring and feedback on a cadence means quality problems are found by the team rather than reported by users. Refreshing the evaluation set from recent production traffic keeps the yardstick aligned with reality as usage drifts — a static eval slowly stops measuring anything relevant. Gating every proposed change through evaluation before deployment is what allows the system to improve without regressing, which is the balance the phase requires.
Why D is wrong. Waiting for defect reports means users are the monitoring system, and it ignores drift entirely: the world changes even when the code does not, so an untouched system degrades in effectiveness over time.
Why A is wrong. Reflexive adoption of every new model is change without evidence. A model upgrade is a significant variable that should be evaluated against your own task before adoption, since newer does not automatically mean better for your specific workload, cost profile, or latency budget.
Why C is wrong. A calendar-driven prompt rewrite is change for its own sake. It introduces regression risk on a schedule unrelated to any observed problem, and prompts should change in response to evaluation findings, not the month.
Blueprint objective: Support lifecycle phases (discovery, design, handoff, monitoring, iteration).