15 Free Claude Certified Architect – Professional (CCAR-P) Practice Questions with Answers (2026)

A single model call augmented with retrieval is the simplest pattern that meets this need. There is no multi-step path to plan, so an agent adds cost and unpredictability without adding value.

• A is wrong because a multi-agent hierarchy is far more complex than independent Q&A needs.
• B is wrong because answers must come from the handbook, and open web search adds risk of off-policy answers.
• D is wrong because it is costly, hard to keep current as the handbook changes, and does not give section references.

Embeddings capture meaning, not exact strings, so they struggle with codes and identifiers. Lexical search handles exact matches well. Hybrid retrieval serves both query patterns.

• A is wrong because more chunks adds noise and cost without ranking the exact match first.
• B is wrong because a bigger embedding model has the same weakness with arbitrary codes.
• C is wrong because it breaks the natural-language questions that already work.

Compliance is a mix of data minimisation, the right contracts and data handling terms, and access controls with auditability. These are design and governance decisions, not prompt settings.

• B is wrong because a prompt instruction is not a data protection control; the API does not learn from requests in the first place.
• C is wrong because encryption in transit is necessary but nowhere near sufficient.
• D is wrong because you still need audit logs; the goal is to control and protect them, not delete them.

Matching capability to the task keeps quality where it matters and saves money where it does not. The evals per route are what make the trade-off safe. It is also worth measuring a single capable model at lower effort before you build a multi-model router.

• A is wrong because the 15% of complex requests would likely fail.
• C is wrong because it truncates answers instead of reducing unnecessary capability.
• D is wrong because chat users cannot wait up to 24 hours.

This is capability bloat. Loading every tool up front costs context and makes selection harder. Progressive discovery and role-based scoping give the model a small, relevant set at each step.

• B is wrong because fitting more tools in does not make choosing between them easier.
• C is wrong because it removes the structure the model needs to call tools correctly.
• D is wrong because thin descriptions make tool selection worse, not better.

A good eval set reflects the real distribution of work, including the hard cases. Model-based graders are useful at scale, but only after you have checked they agree with human judgement.

• B is wrong because testing on the examples the prompt already contains overstates performance.
• D is wrong because self-grading without a rubric is unreliable and biased toward the model's own output.

An end-to-end design includes the feedback path, not just input, processing and output. Corrections become evaluation data, and changes ship only when they improve on it.

• A is wrong because Claude does not update itself from conversations; improvement has to be engineered.
• B is wrong because it adds cost without targeting the actual errors.
• C is wrong because unstructured email does not reach the eval set or the release process.

Authorisation belongs in the data and tool layer, tied to the real user. If the tool cannot return data the user is not allowed to see, no prompt can talk the model into leaking it.

• A is wrong because prompt rules are not access controls and can be bypassed.
• B is wrong because model size does not change what the service account can read.
• D is wrong because logging detects misuse after the fact; it does not prevent it.

Managed settings sit at the top of Claude Code's settings precedence, and nothing in a developer's own files overrides a managed key. Deny rules there enforce the policy for everyone it is deployed to.

• A is wrong because CLAUDE.md is guidance for the model, not an enforced permission.
• C is wrong because developers control their own user settings and can change them.
• D is wrong because a pre-commit hook runs at commit time and does not stop Claude reading files or running commands.

Prompt reuse through Skills or modular prompt components gives you one source of truth, one owner and version control. An update reaches every team at once, and Skills load only when relevant.

• B is wrong because it keeps the copy-and-paste problem and relies on manual effort.
• C is wrong because it repeats 3,000 words on every turn and still does not fix the drift between teams.
• D is wrong because the model does not know your company's specific brand and compliance rules.

Offline evals reduce risk but cannot capture everything production traffic does. A controlled A/B test with agreed metrics and stop conditions confirms the improvement safely before full rollout.

• A is wrong because it skips live validation for a high-traffic system.
• B is wrong because five hand-picked cases are neither representative nor statistically meaningful.
• D is wrong because a model's opinion of a prompt is not evidence of real-world performance.

Multi-agent designs pay off when work splits into many independent pieces that can run in parallel, each with its own context, and then be combined. Forty separate company investigations fit that shape.

• B is wrong because one well-scoped change needs one agent, or no agent at all.
• C is wrong because sequential dependencies remove the benefit of parallel workers; a workflow fits better.
• D is wrong because a single, fast model call is the right tool for low-latency classification.

High-impact decisions need human-in-the-loop review for adverse outcomes, transparency about reasons, and ongoing measurement of disparities. Fairness is something you measure, not something you instruct.

• A is wrong because consistency does not mean fairness, and full automation removes oversight from high-stakes decisions.
• B is wrong because a prompt line cannot show or guarantee unbiased outcomes.
• C is wrong because bias can come through other features, such as location or industry, so removing names is not enough.

Architects make trade-offs visible with data, recommend a path tied to business value, and get explicit agreement. Writing the decision down helps with handoff and future iterations.

• A is wrong because hidden trade-offs damage trust when they surface later.
• C is wrong because committing to targets your evals say you will miss sets the project up to fail.
• D is wrong because it gives up value now for an uncertain future, without asking the sponsor which goal matters most.