Anthropic CCAR-P (Claude Certified Architect - Professional) Practice Test 8

Question 70

A user submits a report that takes ten minutes. The web connection may close, but the user must retrieve the result later. What contract should the service expose?

  1. A new uncorrelated request whenever status is checked
  2. A durable job identifier with status and result retrieval
  3. An instruction to keep the browser tab open
  4. A synchronous connection with no server-side job record

Explanation

Correct Answer

B:

A job resource decouples processing from one connection. Persisted status and results let the user resume observation without resubmitting the work.

Incorrect Answers

A:

Creating new jobs during polling duplicates processing. Status queries should refer to the existing operation rather than initiate another one.

C:

User behavior cannot guarantee network continuity. The architecture must support retrieval after the connection closes, as the requirement explicitly states.

D:

If the connection ends, the user cannot reliably distinguish a lost response from a lost job. The requested later retrieval needs durable identity.

Question 71

Hundreds of identical read-only analysis requests arrive for the same immutable document version. Which design avoids redundant computation while preserving equivalent results?

  1. Launch a separate full analysis for each waiting caller
  2. Combine requests for different document versions into one result
  3. Share one in-flight computation for the same validated input key
  4. Discard all requests after the first without returning results

Explanation

Correct Answer

C:

Coalescing can let multiple callers await one equivalent operation. The key must include relevant configuration and access scope to avoid mixing non-equivalent requests.

Incorrect Answers

A:

Independent execution preserves isolation but wastes work for this explicitly identical immutable input. It does not address the stated redundancy.

B:

Different versions can contain different facts. Reusing an analysis across them would violate the equivalence needed for safe coalescing.

D:

Dropping callers avoids work by failing their requests. Coalescing instead preserves a response contract for each accepted caller.

Question 72

A nonpreemptive worker can run only one queued report at a time. Report A takes two minutes and expires in three; report B takes two minutes and expires in ten. Both are equally important and arrived together. Which scheduling choice avoids an unnecessary deadline miss?

  1. Run B before A
  2. Choose randomly because the reports have equal business importance
  3. Run A before B
  4. Alternate between the reports every thirty seconds

Explanation

Correct Answer

C:

A has the tighter deadline and finishes in time when scheduled first. B can then finish at minute four, still within its ten-minute limit.

Incorrect Answers

A:

B would finish at minute two, but A would finish at minute four and miss its deadline. Both could have met their deadlines with the opposite order.

B:

Equal importance does not make the deadlines interchangeable. Random ordering introduces a preventable miss for A in one of the possible sequences.

D:

The worker is explicitly nonpreemptive, so this schedule is unavailable. A valid choice must respect the execution contract as well as the deadlines.

Question 73

A stateless web tier sends successive turns to different servers. Conversation history currently exists only in each server's memory. What should change?

  1. Store every session under one global conversation key
  2. Keep state only on the worker that handled the previous turn
  3. Persist conversation state behind an authenticated session identifier
  4. Increase the idle timeout of the first server

Explanation

Correct Answer

C:

Shared durable state lets any authorized server reconstruct the correct conversation. Local process memory cannot support continuity across arbitrary routing changes.

Incorrect Answers

A:

A global key mixes separate conversations and cannot reliably associate state with the authenticated user. The backend needs session-scoped identity and access.

B:

Worker-local state can disappear or become unavailable when a later turn reaches another worker. The scenario requires continuity beyond that process.

D:

A longer timeout does not ensure that later turns reach that process. It leaves the fundamental state-placement dependency unresolved.

Question 74

A document pipeline spends more time coordinating hundreds of tiny agents than analyzing pages. Each task does trivial work. Which adjustment should be tested?

  1. Split each page task into more independent agents
  2. Reduce only the length of the final report
  3. Group related low-cost work into coarser tasks
  4. Create another supervisor for every tiny task

Explanation

Correct Answer

C:

Delegation has overhead in planning, messaging, and synthesis. Coarser boundaries can preserve useful parallelism while reducing coordination that dominates task execution.

Incorrect Answers

A:

Further fragmentation makes the same imbalance worse. Parallelism is not beneficial when orchestration costs exceed the work it distributes.

B:

The observed cost occurs during coordination, before final reporting. Output trimming does not directly remove the excessive task-management work.

D:

Additional supervision increases the overhead already causing the problem. It needs a quality benefit that outweighs the added coordination cost.

Question 75

A coordinator assigns a collection of documents to workers, but retries sometimes leave one document unassigned while another is reviewed twice. What should the orchestration layer track?

  1. The total number of worker responses without document identifiers
  2. The average quality score of the documents that received a review
  3. A manifest of required document identities and terminal task outcomes
  4. The combined length of all worker summaries as a completion estimate

Explanation

Correct Answer

C:

A coverage manifest permits reconciliation between required work and observed completion. It exposes both omissions and repeated processing independently of the prose summaries.

Incorrect Answers

A:

A duplicate response can make the count appear complete while a document is missing. Coverage requires correspondence to the actual required identities.

B:

A high score for completed work says nothing about omitted documents. The coordinator must account for the full required collection.

D:

Summary length can vary by document and include duplicate work. It is not a reliable measure of which required documents reached a terminal outcome.

Question 76

An agent's plan says a customer record was updated, but the tool returned a failure. What state should downstream steps consume?

  1. A success flag whenever the tool was invoked
  2. The plan because it was generated before execution
  3. The last successful value without recording the failed update
  4. The observed tool outcome, with the update marked incomplete

Explanation

Correct Answer

D:

A plan describes intended work, while execution evidence establishes what happened. Downstream processing must not assume an unsuccessful operation committed.

Incorrect Answers

A:

Invocation and completion are different events. A call can fail validation, authorization, transport, or execution without changing the target record.

B:

Earlier intent cannot override later evidence of failure. Treating the plan as fact would let the workflow continue on a false premise.

C:

The last known state may still be useful, but the attempted change must not disappear from workflow status. Downstream steps need to know that this update failed rather than assume current completion.

Question 77

An agent discovers that a required document does not exist. Repeating the same retrieval will not change that fact. Which response best repairs the plan?

  1. Mark the document as reviewed to preserve progress
  2. Revise the dependent steps or escalate the missing prerequisite
  3. Continue dependent analysis with an invented document summary
  4. Retry the identical retrieval until the budget expires

Explanation

Correct Answer

B:

The failure changes what can be accomplished. A useful planner updates its dependencies and chooses an allowed alternative instead of repeating an impossible step.

Incorrect Answers

A:

A fabricated completion invalidates downstream conclusions. Progress tracking must reflect what evidence was actually obtained and examined.

C:

Supplying invented evidence conceals the blockage. The plan should expose the dependency failure rather than manufacture the input needed to proceed.

D:

Retries help transient failures, not a confirmed missing prerequisite. This would consume the remaining budget without creating the absent document.

Question 78

A claims assistant sometimes jumps from submitted directly to paid, skipping a required reviewed state. Which design directly enforces valid transitions?

  1. A dashboard showing skipped states after payment
  2. An application state machine that validates transition requests
  3. A model vote on whether the shortcut is acceptable
  4. A longer narrative describing the expected sequence

Explanation

Correct Answer

B:

Explicit transition rules prevent prohibited state changes regardless of the generated suggestion. The model can request an action without owning the authority to commit it.

Incorrect Answers

A:

Observation helps investigation but occurs after the invalid transition. The requirement calls for prevention at the point of state change.

C:

A required business transition is not subject to probabilistic consensus. Voting would replace an enforceable rule with an uncertain judgment.

D:

Narrative guidance may improve adherence but cannot guarantee transition enforcement. The state-changing service must reject invalid moves.

Question 79

A reconciliation agent alternates between two proposed classifications without gaining evidence. Which termination rule best detects this failure?

  1. Double the budget after every repeated classification
  2. Continue whenever the explanation wording changes
  3. Stop only when the model claims complete certainty
  4. Stop on repeated state without information gain

Explanation

Correct Answer

D:

A loop can consume resources while revisiting the same alternatives. Detecting repeated state and missing progress supports escalation before the full budget is wasted.

Incorrect Answers

A:

Increasing resources rewards the detected cycle without addressing its cause. A no-progress condition should trigger repair or a controlled stop.

B:

Different wording does not establish new evidence or progress. A loop can paraphrase the same unresolved choice indefinitely.

C:

Self-reported certainty is not a reliable convergence signal. The agent may remain uncertain indefinitely or become confident without additional support.

Question 80

A design sends 80% of daily cases to reviewers, but their capacity covers only 20%. What architectural issue must be resolved before rollout?

  1. The escalation policy and staffing capacity are incompatible
  2. Set a daily review cap and automatically approve all overflow cases
  3. Keep the escalation rate and size the service only for model throughput
  4. The queue can grow without affecting service objectives

Explanation

Correct Answer

A:

Human oversight is a real processing dependency with finite throughput. The design must reduce eligible volume safely, increase capacity, or narrow the rollout.

Incorrect Answers

B:

A capacity cap does not establish that overflow is safe to approve. The escalation policy must be reconciled with staffing and permitted fallback behavior.

C:

Model capacity does not supply the missing reviewer capacity. A design that ignores the human stage can meet inference targets while failing the end-to-end service objective.

D:

An unbounded queue eventually creates unacceptable delay and stale cases. Queueing cannot compensate indefinitely for a sustained capacity deficit.

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