Anthropic CCAR-P (Claude Certified Architect - Professional) Practice Test 1

Question 1

An insurer already has a fixed sequence for extracting claim fields, checking completeness, and drafting a summary. Each stage has a testable output contract. The team needs reproducible execution and stage-level retries. Which design best fits?

  1. Independent agents that each execute the entire claim process
  2. A checkpointed workflow with validation and separate retries between defined stages
  3. A general agent that chooses and records its own execution path
  4. One model call followed by a retry of the entire claim on failure

Explanation

Correct Answer

B:

Application code can preserve the required order and validate each intermediate artifact. A failed extraction or completeness check can then be retried without discarding a valid summary input or repeating unrelated work.

Incorrect Answers

A:

Duplicating the full process creates reconciliation work rather than useful specialization. It also multiplies processing cost without providing the requested stage-level control.

C:

Allowing the model to decide the process introduces variation where the business has already specified the sequence. That flexibility makes a failed stage harder to isolate and restart.

D:

A combined call may be simple, but the absence of intermediate validation removes the retry boundaries required by this design. A malformed extraction could silently contaminate the later summary.

Question 2

A research assistant investigates unfamiliar supplier issues. The useful next source depends on what earlier searches reveal, and the number of searches cannot be predetermined. Tools are read-only and a maximum investigation budget is enforced. Which approach is most appropriate?

  1. Choose an investigation plan once, then execute it without adapting to findings
  2. Run a fixed industry checklist with the same source sequence for every supplier
  3. Choose follow-up tools adaptively within the budget, using explicit stopping criteria
  4. Allocate the budget equally across sources before inspecting their relevance

Explanation

Correct Answer

C:

The information dependency makes a flexible investigation loop useful. Read-only access, a budget ceiling, and a defined completion condition bound the additional autonomy without forcing a rigid path.

Incorrect Answers

A:

Planning once can bound execution, but it prevents later evidence from changing the investigation path. This task specifically requires follow-up decisions based on intermediate discoveries.

B:

A fixed search plan works when the required sources are known in advance. Here it could miss the source revealed by an unexpected earlier finding.

D:

Equal allocation is predictable but treats sources as equally valuable before any evidence is available. Adaptive investigation can spend the remaining budget on the lead that earlier results support.

Question 3

A procurement review needs security, delivery-capacity, and contract-exclusion assessments. Each specialist reads different documents; the work is independent, and capacity permits all three reviews concurrently. A final recommendation must address all three. Which plan best reduces elapsed time?

  1. Run the three specialist reviews sequentially and combine their structured findings
  2. Run the specialist reviews concurrently, then reconcile their structured evidence and findings
  3. Ask one reviewer to process the complete combined document collection sequentially
  4. Run the reviews concurrently and publish the first completed recommendation

Explanation

Correct Answer

B:

Independent analysis can proceed at the same time, after which a synthesis step resolves contradictions and cites the supporting findings. Explicit output contracts keep concurrency from producing incompatible reports.

Incorrect Answers

A:

Serial review can preserve specialist quality, but it puts independent work on the critical path. The scenario explicitly permits concurrency without waiting for another assessment.

C:

One reviewer can consider all the evidence, but sequential processing does not exploit independent work streams. It does not meet the requested reduction through decomposition.

D:

A race returns quickly by discarding unfinished assessments. The recommendation requires all three perspectives, so the first completion is not an adequate synthesis.

Question 4

A contract pipeline extracts fields, selects a governing jurisdiction, and analyzes the corresponding policy corpus. A reviewer corrects the jurisdiction after compliance analysis; extracted fields remain valid. Which recovery plan restores consistency with the least unnecessary reprocessing?

  1. Keep the old compliance result and append the corrected jurisdiction to its summary
  2. Rerun every stage, including unrelated document extraction that has not changed
  3. Invalidate the dependent compliance result and rerun it using the corrected jurisdiction
  4. Reuse the cached compliance result because the contract identifier is unchanged

Explanation

Correct Answer

C:

The compliance output depends on the jurisdiction decision and must be recomputed when that input changes. Unaffected extraction can be retained if its inputs and relevant versions remain unchanged.

Incorrect Answers

A:

The original analysis used a different policy corpus. Appending a corrected label cannot make its conclusions valid under the newly selected jurisdiction.

B:

A complete restart could produce a valid answer, but it repeats independent work the scenario says is still valid. Dependency-aware invalidation can achieve consistency with less reprocessing.

D:

The contract identifier alone does not capture the inputs used in the analysis. A cache entry must account for the jurisdiction and applicable policy version to avoid stale conclusions.

Question 5

A Claude assistant can propose changes to a production database schema. Analysts want rapid exploration, but applying a migration requires an accountable engineer to approve the exact change. Where should the approval gate be placed?

  1. Require approval before exploration begins, using the requested objective as the artifact
  2. Require approval after generating the SQL but after the execution service has applied it
  3. Require approval of the model-generated plan before the final SQL has been produced
  4. Require engineer approval of the generated SQL before granting production execution permission

Explanation

Correct Answer

D:

This gate leaves read-only exploration fast while placing accountable review before a production side effect. The engineer can inspect the proposed SQL, its scope, and the deployment implications before authorizing execution.

Incorrect Answers

A:

Approving a general objective does not let the engineer review the concrete migration. Exploration may change which SQL is needed after that initial decision.

B:

Post-execution review can detect a bad migration but cannot satisfy a requirement for prior authorization. Some consequences may already require costly recovery.

C:

A plan describes intended behavior, but implementation can introduce additional changes or mistakes. The engineer needs to inspect the final SQL at the action boundary.

Question 6

A service desk prototype reduces response-writing time by 40%, but resolving tickets still takes the same total staff time because agents spend longer checking inaccurate suggestions. Which measure should determine whether the prototype delivers its intended efficiency benefit?

  1. Mean first-response time, excluding time spent checking suggestions
  2. Drafts accepted without editing, measured as a share of generated suggestions
  3. Model processing cost per ticket, excluding staff verification costs
  4. Total staff time per resolved ticket at the required quality level

Explanation

Correct Answer

D:

The business benefit depends on total labor required to complete a ticket correctly. Including verification work and a quality threshold prevents a local drafting improvement from being mistaken for overall efficiency.

Incorrect Answers

A:

A faster initial response may be useful, but it can coexist with slower checking and resolution. It omits the labor that offset the drafting improvement in this scenario.

B:

Acceptance without editing measures one aspect of usability, but staff may still spend substantial time checking accepted drafts. It does not measure total labor saved.

C:

Model expenditure matters to total cost, but the stated efficiency problem is unchanged staff effort. Measuring only compute expense would miss that trade-off.

Question 7

An internal assistant combines inventory results with a carrier tracking service. Carrier requests sometimes time out. Inventory-only answers remain useful if clearly labeled, and no delivery-date commitment may be invented. Which fallback best fits?

  1. Use bounded retries, then answer with the most recent cached delivery estimate regardless of age
  2. Use bounded retries, then return an unavailable response for the entire combined request
  3. Use bounded retries, then predict a delivery date from typical carrier transit times
  4. Return the verified inventory result while clearly identifying tracking information as unavailable

Explanation

Correct Answer

D:

The response preserves the usable part of the result while exposing its limitation. The application should enforce a timeout and prevent the missing tracking field from becoming an inferred promise.

Incorrect Answers

A:

A cached estimate might support a qualified answer under a defined freshness policy. Here no such evidence or policy establishes that an arbitrary old estimate is valid for the current request.

B:

Failing the entire response contains the carrier failure but also discards usable inventory evidence. The business explicitly accepts a clearly identified partial answer.

C:

A transit-time prediction is not a verified tracking result. It would create the unsupported commitment that the fallback must avoid.

Question 8

Two Claude configurations satisfy all acceptance criteria on a representative evaluation, including at least 97% extraction accuracy and the required latency. X costs less; Y is slower with no demonstrated benefit on this workload. The goal is lowest cost while maintaining the criteria. Which deployment is best justified?

  1. Deploy Y for every invoice, reserving X for a later cost-reduction exercise
  2. Run X and Y on every invoice and reconcile their extraction results
  3. Deploy X and monitor production extraction quality and latency against the agreed criteria
  4. Route every invoice to Y unless X can prove the extraction is error-free

Explanation

Correct Answer

C:

The decision follows measured task performance and the cost-latency requirements. Monitoring remains necessary because production inputs can differ from the evaluation set.

Incorrect Answers

A:

The measurements do not show an outcome that justifies the extra expense or delay. Deferring the economical configuration does not satisfy the stated optimization objective.

B:

Dual execution can be useful when verified complementary errors justify an ensemble. No such benefit is established here, and the extra calls increase cost for an already accepted workload.

D:

Requiring certainty from X adds a stricter condition than the agreed acceptance criteria. It can effectively route everything to Y without evidence that this improves the required outcome.

Question 9

A ticket classifier follows the requested JSON shape but repeatedly confuses two categories whose boundary depends on business-specific examples. Which prompt change most directly addresses that confusion?

  1. Add examples of easy categories to reinforce the existing output schema
  2. Add contrasting boundary examples with validated labels and explicit category decision criteria
  3. Expand the category descriptions using synonyms but retain the same examples
  4. Have the model generate its own labeled examples without business review

Explanation

Correct Answer

B:

Examples near the ambiguous boundary demonstrate how the stated rules apply to realistic tickets. Including contrasting cases is more informative than repeating easy examples of already distinct categories.

Incorrect Answers

A:

Additional easy cases can reinforce familiar behavior but do not clarify the disputed distinction. The schema is already correct, so format demonstrations target the wrong symptom.

C:

More descriptive wording may help readability, but synonyms do not necessarily show how the business resolves the ambiguous boundary. Contrasting validated examples address that gap directly.

D:

Self-generated labels can reproduce the same misunderstanding. Examples meant to establish a business rule need to reflect the actual approved decisions.

Question 10

An engineering assistant receives thousands of lines of obsolete build logs in every request. Recent error evidence and relevant repository instructions are frequently omitted to fit the context limit. Which change should be tried first?

  1. Keep an equal token allowance from every historical build, truncating each proportionally
  2. Summarize only the old build logs and continue omitting current repository instructions
  3. Prioritize current diagnostic evidence and relevant instructions; retrieve older logs when needed
  4. Retain error-frequency counts from all builds and discard individual diagnostic messages

Explanation

Correct Answer

C:

The immediate failure is poor selection of information, not an inherent need to send every available record. Keeping high-value current evidence makes the limited input more useful for diagnosis.

Incorrect Answers

A:

Equal allocation treats stale and current evidence as equally valuable. It can still truncate the recent messages and relevant instructions that are needed for this failure.

B:

Compression can save space, but this selection still excludes information directly relevant to the current environment. Summarization is not a substitute for choosing the right evidence.

D:

Counts can identify recurring symptoms but discard stack traces, locations, and context. The assistant needs specific diagnostic evidence, not only an aggregate frequency report.

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