Anthropic CCA-F (Claude Certified Architect - Foundations (CCAR-F)) Practice Test 5

Question 40

A policy assistant cites the correct current policy, but an answer sentence attributes a benefit to that policy that its text does not support. Which review step directly detects and corrects this attribution error?

  1. Verify that the cited URL resolves, then accept every sentence attached to it.
  2. Confirm that the policy title contains the same keywords as the benefit.
  3. Add more citations about similar benefits from other organizations.
  4. Compare the sentence with the cited passages and remove or correct the unsupported attribution.

Explanation

Correct Answer

D:

A claim-level support check tests whether the evidence actually establishes the statement. If the policy does not support it, the answer should not present it as that policy's rule.

Incorrect Answers

A:

A working link proves source availability, not that the source supports a particular claim. The cited text must be checked against the assertion.

B:

Topical relevance is weaker than support for the claim. A policy can mention a benefit category without granting the specific entitlement stated in the answer.

C:

Other organizations' rules do not substantiate this policy attribution. More references cannot repair the mismatch between the stated source and the asserted entitlement.

Question 41

A client requests structured JSON from a supported Claude model. Which TWO returned stop reasons require special handling because the response may not contain the complete requested schema-conforming object? Choose TWO.

  1. max_tokens, with the response cut off at the output limit.
  2. refusal, with a response declining the request.
  3. end_turn, with required properties returned in a different order.
  4. end_turn, with a field whose null value is explicitly allowed by the schema.
  5. end_turn, with a complete response that passes the application's schema check.

Explanation

Correct Answers

A:

The output budget can interrupt generation before the object is complete. The caller must detect truncation and recover rather than passing an unfinished result to its consumer.

B:

A refusal can take precedence over the requested schema. The client needs a refusal branch instead of assuming the content is an ordinary extraction record.

Incorrect Answers

C:

JSON object property order does not change the meaning of the object for schema validation. A consumer should validate named properties rather than assume a fixed textual order.

D:

A permitted null can intentionally represent unavailable data. It is a valid value under the stated contract, not a stop condition that overrides the requested structure.

E:

A normal end of turn with validated output is the intended structural success path. Business-value checks may still be needed, but this condition does not itself indicate incomplete generation.

Question 42

A completed Message Batch contains successful document extractions, invalid-request errors, and expired requests. Successful records have already passed validation. Which THREE actions support recovery without rerunning those successful records? Choose THREE.

  1. Select the expired requests for retry under the remaining application deadline.
  2. Correct invalid-request causes before resubmitting the affected requests.
  3. Match each returned outcome to its input using custom_id.
  4. Treat processing_status ended as evidence that every request succeeded.
  5. Join returned results to input documents by line number in the results file.
  6. Resubmit every original request to preserve its original position in the batch.

Explanation

Correct Answers

A:

An expired request did not produce a completed model result in that batch. It can be reconsidered for retry without discarding records that already succeeded.

B:

A malformed request is not repaired by repeating it unchanged. Fixing the identified input problem makes a targeted retry meaningful.

C:

Batch result ordering need not match submission order. A stable request identifier allows the recovery process to locate the correct source document and its outcome.

Incorrect Answers

D:

The batch can finish with mixed outcomes. Its lifecycle status must not replace inspection of the individual result types.

E:

Ordering is not guaranteed, so line-based association can attach the wrong result to a document. Use the per-request identifier instead.

F:

Reprocessing validated successes is unnecessary under the stated goal. Position is not the mechanism for correlating batch results.

Question 43

A product assistant retrieves webpages that may contain instructions to disclose secrets. Which TWO architectural controls directly address the untrusted-content boundary and the possible exposure of credentials? Choose TWO.

  1. Keep credentials outside model-visible task data and enforce tool-side restrictions on secret access.
  2. Promote retrieved text into the system prompt after checking that the page uses HTTPS.
  3. Allow secret reads when a retrieved page claims the operation is part of its verification procedure.
  4. Treat the sentence as untrusted page content rather than an authorized instruction.
  5. Remove only the phrase reveal your API key and trust the remaining page text as instructions.

Explanation

Correct Answers

A:

Limiting exposure reduces what an injected instruction could obtain, while execution controls restrict available actions. This complements instruction-level handling rather than assuming perfect resistance from the model alone.

D:

Retrieval supplies evidence for the product question; it does not delegate control of the assistant. The malicious sentence has no authority to change the task or request secret disclosure.

Incorrect Answers

B:

HTTPS protects transport to the server; it does not make that server's content authoritative instructions. Promotion would grant external text inappropriate influence.

C:

An external page cannot authorize access to application secrets. A claim of necessity within that same untrusted content does not change its authority.

E:

A literal phrase filter misses paraphrases and other injected commands. It does not establish a general trust boundary for external content.

Question 44

A long-running agent must compact its history. Which information is most important to preserve for correct continuation of an unfinished migration?

  1. Confirmed decisions, current state, unresolved failures, constraints, and references to necessary artifacts.
  2. Preserve completed steps but omit unresolved failures so the summary stays concise.
  3. Keep the most verbose tool outputs and discard decisions that lack a recent timestamp.
  4. Retain recent messages verbatim and discard the original migration requirements without extracting them.

Explanation

Correct Answer

A:

Continuation depends on knowing what has happened and what still governs the work. Artifact references allow the agent to retrieve detailed evidence without retaining every raw observation in context.

Incorrect Answers

B:

The next stage needs to know which problems remain open. Omitting them can cause a failed step to be treated as complete or lead the agent to repeat an unsuccessful approach.

C:

Verbosity and recency are weak substitutes for task relevance. An older architectural constraint can still govern the migration after recent diagnostic output becomes obsolete.

D:

A rolling window can lose constraints that are no longer repeated. Compaction should preserve their current meaning even when the original wording is removed.

Question 45

A research task spans sessions whose context may be reset. The agent must retain which sources were checked and which claims remain unresolved. Which design best supports continuity?

  1. Save only the final prose answer, omitting which sources were checked and which claims remain open.
  2. Persist the raw tool logs but provide the resuming agent no index or retrieval path to them.
  3. Keep the progress record only in the current process's in-memory conversation buffer.
  4. Maintain structured persistent notes with source references, decisions, and open questions.

Explanation

Correct Answer

D:

A durable record can be reloaded after transient context disappears. Keeping provenance and unresolved status prevents the next session from mistaking an incomplete investigation for a verified conclusion.

Incorrect Answers

A:

A final narrative can conceal unfinished verification and rejected hypotheses. Continuation requires explicit progress and provenance, not just a polished output.

B:

Storage alone is insufficient if the next session cannot locate relevant state. Structured notes and references connect the durable evidence to the unfinished work.

C:

A buffer that disappears at reset cannot support the next session. The application needs a persistent record and a way to load the relevant entries.

Question 46

A policy assistant retrieves an obsolete policy with excellent keyword overlap and a current policy with slightly lower overlap. The user asks about today's entitlement. Which retrieval decision best supports accuracy?

  1. Use effective dates and supersession metadata alongside relevance when selecting evidence.
  2. Merge entitlements from both versions to maximize the answer's coverage.
  3. Prefer whichever policy the crawler indexed most recently, regardless of its effective date.
  4. Select only by keyword similarity, ignoring effective dates.

Explanation

Correct Answer

A:

The requested answer depends on the policy in force today. Version-aware retrieval helps ensure that the source matches the user's temporal scope before the model synthesizes it.

Incorrect Answers

B:

The older version may contain superseded provisions. Combining grants without resolving applicability can invent a policy that was never actually in force.

C:

A fresh crawl can index an obsolete document. Retrieval freshness and policy validity are different attributes, so the applicable effective period must be checked.

D:

Lexical overlap measures relevance to wording, not temporal applicability. An obsolete document can closely match the question while giving the wrong current rule.

Question 47

Requests share a long static policy followed by a timestamp and a different user question. The only explicit cache breakpoint is after the changing question. No earlier prefix has been cached, and cache usage shows repeated writes without reads. What change most directly enables reuse of the static policy?

  1. Move the timestamp before the policy and leave the breakpoint after the question.
  2. Place an explicit breakpoint at the end of the unchanged policy, before the per-request material.
  3. Increase the response token budget while leaving the prompt arrangement unchanged.
  4. Increase the cache lifetime while leaving the only breakpoint after the changing question.

Explanation

Correct Answer

B:

The reusable entry then covers a stable prefix. Later questions and timestamps can change without changing the policy entry that subsequent requests can read.

Incorrect Answers

A:

The changing timestamp still lies inside the cached prefix. Reordering it earlier does not create a stable cache entry.

C:

The response budget controls generated output. It does not change the identity of the input prefix used for cache matching.

D:

A longer lifetime does not make different prefixes identical. The varying suffix continues to change the prefix represented by that breakpoint.

Question 48

A read-only API request fails with a documented temporary overload error. The application has a bounded retry budget. Which recovery policy is most appropriate?

  1. Retry with bounded backoff and jitter, then report a controlled failure if the budget is exhausted.
  2. Retry indefinitely whenever the request is read-only, because it cannot create duplicate writes.
  3. Retry at the maximum client concurrency with no delay between attempts.
  4. Apply the same long fixed sleep to every client before each retry.

Explanation

Correct Answer

A:

Backoff spreads attempts while bounded retries prevent indefinite waiting. A clear exhausted-budget path preserves honest status when the dependency remains unavailable.

Incorrect Answers

B:

Read-only status avoids one class of side-effect risk but does not justify unlimited resource use or user waiting. Recovery still needs a bounded failure path.

C:

Immediate concurrent retries can amplify overload and synchronize failures. They consume the retry budget without giving the dependency time to recover.

D:

A fixed delay can reduce frequency, but synchronized clients may retry together and create another burst. Jitter and an explicit budget better control this behavior.

Question 49

An assistant answers 95% of all test cases correctly, but nearly every case involving a missing document fails. Those cases are rare in the aggregate set. What should the team do before concluding that the assistant is ready?

  1. Inspect performance by failure category and evaluate missing-document behavior explicitly.
  2. Set deployment approval solely from the 95% aggregate score because the missing-document group is small.
  3. Increase the number of ordinary cases until the aggregate score has a narrower confidence interval.
  4. Measure only response latency for missing-document cases because their factual answers cannot be evaluated.

Explanation

Correct Answer

A:

Aggregate accuracy can conceal severe weaknesses in small but important groups. Category-level evaluation reveals whether the system handles absence of evidence safely instead of inventing an answer.

Incorrect Answers

B:

A rare category may still carry substantial operational risk. Its acceptable failure rate must be assessed explicitly rather than inferred from its frequency.

C:

That can improve precision for the overall estimate while leaving the rare failure mode unchanged. It does not establish acceptable performance for missing-document requests.

D:

The absence of a document still has an evaluable expected behavior, such as acknowledging missing evidence. Fast unsupported answers do not satisfy that requirement.

Question 50

An extraction assistant assigns confidence 0.95 to many records, but human review finds that only about 70% of those records are correct. The team wants to use confidence to route risky records for review. What should it do first?

  1. Reduce review volume by accepting every record above the current 0.90 threshold.
  2. Interpret 0.95 as a guaranteed 95% correctness probability for each record.
  3. Measure score reliability on representative labeled cases and choose review thresholds from observed error rates.
  4. Average several confidence statements from the same generation and use their mean as ground truth.

Explanation

Correct Answer

C:

Calibration relates reported confidence to actual outcomes. Thresholds should then reflect the cost of errors and review capacity, with continued monitoring for distribution changes.

Incorrect Answers

A:

The current scores overstate reliability in the described group. Applying a lower unvalidated cutoff would not establish an acceptable risk level.

B:

The observed error rate contradicts that interpretation. A model-reported score is not automatically calibrated to empirical correctness.

D:

Repeated self-assessments can share the same bias. Averaging them does not provide the labeled outcomes needed to measure reliability.

Claude Certified Architect - Foundations (CCAR-F) (CCA-F) Practice Tests