Anthropic CCAR-P (Claude Certified Architect - Professional) Practice Test 2

Question 10

A team assumes a proof of concept can write directly into a billing system because reviewers liked the generated text. What should the solution architect prioritize?

  1. Treat grounded evidence and reviewability as first-class product requirements.
  2. Define a business KPI, baseline process cost, and acceptable error budget before selecting an architecture.
  3. Distinguish informational output from authenticated actions in the business capability map.
  4. Separate low-risk drafting from a gated closure workflow with explicit authorization and human sign-off.

Explanation

Correct Answer

C:

Drafting and financial posting have different failure consequences; the architecture must identify who may commit each side effect.

Incorrect Answers

A:

Improving traceable support for generated claims does not imply transaction authorization requirements. The selected design must distinguish informational output from authenticated actions in the business capability map, whereas this choice omits that direct protection or verification step.

B:

This addresses measurable business acceptance criteria; however, the problem asks for transaction authorization requirements. Without the safeguard to distinguish informational output from authenticated actions in the business capability map, the original failure mode is still possible.

D:

This would make sense as an additional project activity for a safe boundary around irreversible actions. The priority required here is transaction authorization requirements, so the workflow still needs a control that will distinguish informational output from authenticated actions in the business capability map.

Question 11

A manufacturer expects a knowledge assistant to answer policy questions but its source files have unclear ownership and no update process. Which design choice best matches the stated constraint?

  1. Classify use cases and data sensitivity before choosing ingestion, deployment, and retention controls.
  2. Treat grounded evidence and reviewability as first-class product requirements.
  3. Name content owners and freshness SLAs as prerequisites for an answer architecture.
  4. Define a business KPI, baseline process cost, and acceptable error budget before selecting an architecture.

Explanation

Correct Answer

C:

Retrieval quality depends on governed source material; stale manuals cannot be repaired by a larger language model alone.

Incorrect Answers

A:

The option belongs to the broader work on data classification driving architecture, but it cannot by itself deliver ownership and freshness of source knowledge. A decision on this case needs the team to name content owners and freshness SLAs as prerequisites for an answer architecture and verify that outcome.

B:

This shifts attention toward traceable support for generated claims when the relevant decision is about ownership and freshness of source knowledge. Unless the architecture can name content owners and freshness SLAs as prerequisites for an answer architecture, the failure described in the stem remains unresolved.

D:

Although measurable business acceptance criteria can be valuable, it is not the acceptance condition in the question. The missing provision is ownership and freshness of source knowledge, addressed by the control to name content owners and freshness SLAs as prerequisites for an answer architecture.

Question 12

A public agency requests automated citizen decisions without establishing legal appeal or transparency obligations. Which architecture adjustment is most appropriate?

  1. Add appealability, explainability and accountability to the use-case acceptance requirements.
  2. Collect representative locale-specific samples and acceptance criteria during discovery.
  3. Treat grounded evidence and reviewability as first-class product requirements.
  4. Classify use cases and data sensitivity before choosing ingestion, deployment, and retention controls.

Explanation

Correct Answer

A:

Regulated decision impact changes the feasible operating model; the system should not be designed solely around throughput.

Incorrect Answers

B:

There is a legitimate use case for language-specific requirements and coverage, but it does not supply the test or enforcement needed for accountable decisions and appeal processes. The design needs to add appealability, explainability and accountability to the use-case acceptance requirements.

C:

Consider its actual operating scope: traceable support for generated claims. The requested assurance is accountable decisions and appeal processes, which calls for an implementation that can add appealability, explainability and accountability to the use-case acceptance requirements, not merely for a related improvement.

D:

The alternative could reduce risk associated with data classification driving architecture, yet the described behavior requires accountable decisions and appeal processes. That outcome depends on the capability to add appealability, explainability and accountability to the use-case acceptance requirements.

Question 13

A compliance intake service receives scans, retrieves policy, proposes classifications, and posts cases to a queue; failures have no traceable source. Which next step is most defensible?

  1. Persist workflow checkpoints and deduplicate external actions using transaction IDs.
  2. Define explicit ingestion, extraction, retrieval, inference, validation, human review and case-posting stages with correlation IDs.
  3. Capture per-stage evidence, latencies, model identifiers, retrieved document IDs and validation outcomes.
  4. Place jobs on a durable queue with workers, state tracking, deadlines and retry budgets.

Explanation

Correct Answer

B:

A stage boundary makes failures diagnosable and supports replay without repeating authorized side effects. Place a verifiable contract or checkpoint at the stage responsible for traceable stage-by-stage execution, then test recovery without silently repeating external writes.

Incorrect Answers

A:

This would make sense as an additional project activity for recovery without repeating committed actions. The priority required here is traceable stage-by-stage execution, so the workflow still needs a control that will define explicit ingestion, extraction, retrieval, inference, validation, human review and case-posting stages with correlation IDs.

C:

Its main effect concerns observability across pipeline stages. The case explicitly hinges on traceable stage-by-stage execution; to resolve it reliably, the implementation has to define explicit ingestion, extraction, retrieval, inference, validation, human review and case-posting stages with correlation IDs, which this choice does not establish.

D:

The option might improve durable asynchronous job processing; that remains insufficient for traceable stage-by-stage execution. The case is resolved by traceable stage-by-stage execution, which depends on a design that will define explicit ingestion, extraction, retrieval, inference, validation, human review and case-posting stages with correlation IDs.

Question 14

A chat summarizer intermittently produces malformed output that crashes a downstream ERP adapter. Which approach should be recommended?

  1. Define explicit ingestion, extraction, retrieval, inference, validation, human review and case-posting stages with correlation IDs.
  2. Validate a typed output contract before invoking the adapter, with retry or exception handling.
  3. Place jobs on a durable queue with workers, state tracking, deadlines and retry budgets.
  4. Use versioned interfaces between parsing, retrieval, generation and response rendering.
  5. Add stateful approval gates and time-bounded publishing authorization.

Explanation

Correct Answer

B:

Structural checks at the application boundary prevent unparseable model text from reaching business APIs. Place a verifiable contract or checkpoint at the stage responsible for machine-enforced output contract, then test recovery without silently repeating external writes.

Incorrect Answers

A:

Although traceable stage-by-stage execution can be valuable, it is not the acceptance condition in the question. The missing provision is machine-enforced output contract, addressed by the control to validate a typed output contract before invoking the adapter, with retry or exception handling.

C:

It emphasizes durable asynchronous job processing without closing the stated gap in machine-enforced output contract. By contrast, the remedy to validate a typed output contract before invoking the adapter, with retry or exception handling acts on the condition that makes this scenario fail.

D:

That is primarily a technique for modular component contracts. It lacks the explicit measure for machine-enforced output contract that is necessary here, namely a system that will validate a typed output contract before invoking the adapter, with retry or exception handling.

E:

Improving time-sensitive workflow approvals does not imply machine-enforced output contract. The selected design must validate a typed output contract before invoking the adapter, with retry or exception handling, whereas this choice omits that direct protection or verification step.

Question 15

A file-enrichment pipeline retries after timeouts and sometimes creates the same invoice twice. Which change most directly satisfies this requirement?

  1. Persist workflow checkpoints and deduplicate external actions using transaction IDs.
  2. Use versioned interfaces between parsing, retrieval, generation and response rendering.
  3. Define explicit ingestion, extraction, retrieval, inference, validation, human review and case-posting stages with correlation IDs.
  4. Use stable request identifiers and idempotent writes at the posting boundary.

Explanation

Correct Answer

D:

Retries are normal in distributed systems; idempotency ensures repeated delivery does not multiply side effects.

Incorrect Answers

A:

The alternative could reduce risk associated with recovery without repeating committed actions, yet the described behavior requires duplicate prevention for retried writes. That outcome depends on the capability to use stable request identifiers and idempotent writes at the posting boundary.

B:

The proposed measure is aimed at modular component contracts, not duplicate prevention for retried writes. The architecture must instead use stable request identifiers and idempotent writes at the posting boundary; a separate change is still needed for duplicate prevention for retried writes.

C:

The relevant trade-off is not primarily traceable stage-by-stage execution. A successful fix needs duplicate prevention for retried writes; otherwise duplicate prevention for retried writes remains unverified even after this change.

Question 16

A research assistant currently drops failed retrievals silently and answers from memory. Which solution best resolves this operational gap?

  1. Propagate retrieval status and enforce a fail-closed or explicit uncertainty response.
  2. Capture per-stage evidence, latencies, model identifiers, retrieved document IDs and validation outcomes.
  3. Define explicit ingestion, extraction, retrieval, inference, validation, human review and case-posting stages with correlation IDs.
  4. Use versioned interfaces between parsing, retrieval, generation and response rendering.

Explanation

Correct Answer

A:

The generation stage needs evidence-health metadata; otherwise a source outage masquerades as confident knowledge.

Incorrect Answers

B:

The proposed action prioritizes observability across pipeline stages. That can coexist with the right solution, but it cannot replace safe degradation when evidence retrieval fails: the application must still propagate retrieval status and enforce a fail-closed or explicit uncertainty response.

C:

The design concern is real but different: traceable stage-by-stage execution. It does not impose the specific requirement for safe degradation when evidence retrieval fails; a suitable solution must propagate retrieval status and enforce a fail-closed or explicit uncertainty response and demonstrate the result.

D:

This addresses modular component contracts; however, the problem asks for safe degradation when evidence retrieval fails. Without the safeguard to propagate retrieval status and enforce a fail-closed or explicit uncertainty response, the original failure mode is still possible.

Question 17

During a case-management rollout, supervisors need disputed AI classifications to become reproducible improvement cases. Separately, the case queue receives out-of-order status events, allowing a stale update to overwrite a later adjudication. Which TWO controls solve these problems? Choose TWO.

  1. Define explicit ingestion, extraction, retrieval, inference, validation, human review and case-posting stages with correlation IDs.
  2. Capture per-stage evidence, latencies, model identifiers, retrieved document IDs and validation outcomes.
  3. Create a feedback channel that links user corrections to trace IDs and a governed evaluation backlog.
  4. Use consented, minimal preference state separate from expiring conversational context.
  5. Reject stale case updates using monotonic revision numbers and compare-and-set checks at the state store.

Explanation

Correct Answers

C:

A correction becomes useful only when it is tied to the original model version, source evidence and trace so reviewers can reproduce it. A governed backlog turns those cases into regression tests rather than allowing disputed labels to be silently fed into production.

E:

A revision guard prevents a delayed event from rolling back newer adjudicated state. Enforce the comparison where case records are written, and retain rejected-event evidence for safe reconciliation rather than trusting event-arrival order.

Incorrect Answers

A:

That is primarily a technique for traceable stage-by-stage execution. It lacks the explicit measure for feedback tied to reproducible cases and prevention of out-of-order case state regression that is necessary here, namely a system that will reject stale case updates using monotonic revision numbers and compare-and-set checks at the state store.

B:

Improving observability across pipeline stages does not imply feedback tied to reproducible cases and prevention of out-of-order case state regression. The selected design must create a feedback channel that links user corrections to trace IDs and a governed evaluation backlog, whereas this choice omits that direct protection or verification step.

D:

This is a valid architectural concern when optimizing minimal persistent memory boundaries. Here it leaves feedback tied to reproducible cases and prevention of out-of-order case state regression unaddressed, whereas the intended change must create a feedback channel that links user corrections to trace IDs and a governed evaluation backlog before the case can be considered resolved.

Question 18

An asynchronous report generator needs to serve tens of thousands of documents without occupying live HTTP requests. Which intervention most directly addresses the failure?

  1. Place jobs on a durable queue with workers, state tracking, deadlines and retry budgets.
  2. Propagate retrieval status and enforce a fail-closed or explicit uncertainty response.
  3. Capture per-stage evidence, latencies, model identifiers, retrieved document IDs and validation outcomes.
  4. Define explicit ingestion, extraction, retrieval, inference, validation, human review and case-posting stages with correlation IDs.

Explanation

Correct Answer

A:

Decoupling ingestion from inference absorbs bursts and makes completion status explicit. Place a verifiable contract or checkpoint at the stage responsible for durable asynchronous job processing, then test recovery without silently repeating external writes.

Incorrect Answers

B:

The option belongs to the broader work on safe degradation when evidence retrieval fails, but it cannot by itself deliver durable asynchronous job processing. A decision on this case needs the team to place jobs on a durable queue with workers, state tracking, deadlines and retry budgets and verify that outcome.

C:

This shifts attention toward observability across pipeline stages when the relevant decision is about durable asynchronous job processing. Unless the architecture can place jobs on a durable queue with workers, state tracking, deadlines and retry budgets, the failure described in the stem remains unresolved.

D:

For a workflow whose primary weakness was traceable stage-by-stage execution, that proposal could be reasonable. Here the unmet obligation is durable asynchronous job processing, and that requires the system to place jobs on a durable queue with workers, state tracking, deadlines and retry budgets.

Question 19

A payment assistant submits a charge, but the request times out before the gateway reports a result. The provider has a separate transaction-status endpoint. What should the recovery process do before declaring failure or issuing another charge?

  1. Retry immediately with a fresh payment identifier because the first response never arrived.
  2. Persist a pending transaction state and reconcile the provider’s authoritative status before deciding whether another attempt is safe.
  3. Mark the charge unsuccessful and close the customer case as soon as any network timeout occurs.
  4. Ask the language model to infer the payment state from the conversation transcript alone.

Explanation

Correct Answer

B:

A client timeout is not proof that the charge was rejected. Querying the provider using the original transaction reference distinguishes committed, failed and pending states; subsequent actions must follow the reconciled result.

Incorrect Answers

A:

The provider may have committed the first charge even though the response was lost. A fresh identifier can cause a separate payment and therefore cannot safely resolve an unknown outcome.

C:

A timeout describes the communication path, not the business transaction. Closing the case without reconciliation can misstate the customer’s balance and leave a completed payment unrecorded.

D:

A transcript cannot supply the payment provider’s committed transaction record. Recovery must obtain authoritative status through a trusted service, not plausibility-based model inference.

Question 20

Operations cannot tell whether low answer quality began in document parsing, retrieval, or the model. Which design choice best matches the stated constraint?

  1. Use stable request identifiers and idempotent writes at the posting boundary.
  2. Capture per-stage evidence, latencies, model identifiers, retrieved document IDs and validation outcomes.
  3. Propagate retrieval status and enforce a fail-closed or explicit uncertainty response.
  4. Define explicit ingestion, extraction, retrieval, inference, validation, human review and case-posting stages with correlation IDs.

Explanation

Correct Answer

B:

End-to-end instrumentation separates ingestion defects from reasoning defects and makes root-cause analysis possible.

Incorrect Answers

A:

This is a valid architectural concern when optimizing duplicate prevention for retried writes. Here it leaves observability across pipeline stages unaddressed, whereas the intended change must capture per-stage evidence, latencies, model identifiers, retrieved document IDs and validation outcomes before the case can be considered resolved.

C:

By itself, this improves safe degradation when evidence retrieval fails, but the case is not asking the architect to optimize that dimension. The requirement is observability across pipeline stages, supported by an approach to capture per-stage evidence, latencies, model identifiers, retrieved document IDs and validation outcomes.

D:

Its main effect concerns traceable stage-by-stage execution. The case explicitly hinges on observability across pipeline stages; to resolve it reliably, the implementation has to capture per-stage evidence, latencies, model identifiers, retrieved document IDs and validation outcomes, which this choice does not establish.

Claude Certified Architect - Professional (CCAR-P) Practice Tests