Anthropic CCA-F (Claude Certified Architect - Foundations (CCAR-F)) Practice Test 40

Question 390

A fixed prompt and dataset produce different outcomes across repeated runs. What should a reliability estimate report?

  1. A guarantee that the next run will match the most recent one.
  2. Change the prompt between every run and attribute all score variation to repeated sampling.
  3. Performance over enough repeated trials to characterize variability on consequential cases.
  4. Only the best run because all runs used the same prompt.

Explanation

Correct Answer

C:

Repeated observations reveal instability that a single favorable run can hide. Reporting the distribution or range of relevant outcomes is more informative than publishing only the best run. Consequential unstable cases may need a stronger control even if average accuracy is acceptable.

Incorrect Answers

A:

Observed variation provides no basis for that deterministic prediction.

B:

Changing the prompt introduces a second cause of variation. Repeated-run reliability needs the relevant configuration held fixed so its variability can be characterized.

D:

Selecting the maximum produces an optimistic estimate of ordinary behavior.

Question 391

A service relays generated tokens to a slow client. Production is faster than delivery, and an unbounded in-memory queue grows until the service crashes. Which design controls this resource failure while making incomplete delivery visible?

  1. Increase the maximum generation length so the producer has more opportunity to finish.
  2. Start another producer whenever the queue crosses a threshold.
  3. Discard arbitrary queued tokens and continue presenting the response as complete.
  4. Bound the buffer and apply supported flow control, with an explicit timeout or cancellation outcome when delivery cannot keep up.

Explanation

Correct Answer

D:

A finite buffer prevents unrestricted memory growth. Flow control addresses the rate mismatch, while a declared terminal outcome avoids disguising an interrupted response as complete.

Incorrect Answers

A:

A larger possible response can create more queued data. It does not address the mismatch between production and client consumption rates.

B:

Additional production increases the backlog when the consumer is already the bottleneck. It cannot drain the existing queue faster.

C:

Dropping tokens corrupts the delivered content. Silent loss would hide truncation instead of providing a reliable resource policy.

Question 392

A queue of independent requests overwhelms the API's rate limits when all workers start together. Which TWO controls are appropriate? Choose TWO.

  1. Use queue backpressure and bounded retry handling for throttled requests.
  2. Apply coordinated concurrency or rate limits across workers.
  3. Hide rate-limit responses from the job controller.
  4. Assign every worker an unlimited retry loop.
  5. Mark queued requests successful before submitting them.

Explanation

Correct Answers

A:

Backpressure slows intake while controlled retry handles temporary rejection without uncontrolled amplification.

B:

Shared admission control prevents individual workers from collectively exceeding the service budget.

Incorrect Answers

C:

Suppressing feedback prevents the controller from adjusting request admission.

D:

Unbounded retries amplify load and can prolong the rate-limited condition.

E:

Queue admission is not evidence that the requests were processed.

Question 393

An API rejects a request because its payload is too large. What is the most appropriate recovery?

  1. Remove all input and fabricate an answer from the file name.
  2. Retry the identical bytes immediately until the service changes its mind.
  3. Increase the response-token limit while keeping the oversized input.
  4. Reduce or partition the input while preserving task boundaries and required evidence.

Explanation

Correct Answer

D:

A size violation requires changing what is submitted, not merely retrying the same request. Partitioning must preserve context needed to interpret each unit and reconcile the complete job afterward. Splitting at arbitrary bytes can satisfy a payload limit while damaging document meaning or syntax.

Incorrect Answers

A:

Eliminating the evidence would no longer perform the requested source-based task.

B:

An unchanged payload remains over the applicable size limit.

C:

Output allowance does not reduce the rejected input payload.

Question 394

An incident timeline combines logs from two services. Event A is stamped 09:15+02:00; Event B is stamped 08:00+00:00 on the same date. The clocks are synchronized, and the offsets are explicit. Which ordering should the timeline use?

  1. Treat the order as unknowable because the events came from different services.
  2. Event B before Event A, because 08:00 is earlier than 09:15 when compared as displayed strings.
  3. Event A before Event B, because A corresponds to 07:15 UTC and B to 08:00 UTC.
  4. Treat the events as simultaneous because their offsets differ.

Explanation

Correct Answer

C:

Normalizing offsets places both timestamps on the same time scale. It shows A occurring 45 minutes earlier despite its later-looking local clock reading.

Incorrect Answers

A:

The scenario supplies synchronized clocks and explicit offsets. Those conditions are sufficient to compare the recorded instants.

B:

That comparison ignores the offsets. The two local clock readings need a common reference before they can establish event order.

D:

Different offsets describe different local representations; they do not imply simultaneous events. Their normalized times are not equal.

Question 395

A research answer must address six requirements, but retrieved evidence supports only four. What should the synthesis stage do?

  1. Use the four retrieved sources to fill all six requirements without checking their coverage.
  2. Map evidence to the six requirements, answer the supported parts, and identify unresolved gaps for further work.
  3. Declare the whole research task impossible and discard the four valid findings.
  4. Write plausible answers for the remaining two without marking the gap.

Explanation

Correct Answer

B:

An explicit coverage inventory prevents omissions from being hidden by a fluent overall narrative. Each requirement should have supported findings or an explicit unresolved status. That lets the workflow request targeted additional evidence without disguising a partial answer as full coverage.

Incorrect Answers

A:

A source can be relevant overall yet provide no evidence for particular required claims.

C:

Partial supported findings can still be useful when their coverage limits are clear.

D:

Unsupported completion conceals the difference between evidence and invention.

Question 396

Human annotators disagree frequently on whether answers meet a reliability rubric. What should happen before using those labels as ground truth?

  1. Choose the label from the fastest annotator in each disputed case.
  2. Remove every disputed example and report agreement only on the remaining easy cases.
  3. Average the conflicting labels into a score before clarifying what each label is meant to measure.
  4. Clarify ambiguous criteria, adjudicate representative disagreements, and measure agreement again.

Explanation

Correct Answer

D:

A consistent labeling standard is needed before scores can meaningfully evaluate model behavior. Adjudication should explain which rubric interpretation applies to disputed examples. If humans cannot apply the rule consistently, a model score against those labels will be difficult to interpret reliably.

Incorrect Answers

A:

Annotation speed does not establish that the rubric was applied correctly. This rule hides disagreement without improving the criteria or reference labels.

B:

Filtering out disagreement can inflate the agreement measure while leaving the hard decision boundaries undefined. The intended evaluation population would no longer be represented.

C:

Aggregation cannot repair inconsistent interpretations of the criterion. A numeric result would conceal ambiguity in the supposed ground truth.

Question 397

An external contributor adds a false policy page that ranks highly in retrieval. Which TWO controls address this at the knowledge-source layer? Choose TWO.

  1. Promote the page to authoritative status solely because its retrieval score is high.
  2. Lower the confidence score in answers while leaving the false page authoritative in the index.
  3. Track source ownership and provenance, with appropriate review before authoritative inclusion.
  4. Treat multiple mirrors of the contributed page as independent approval for admitting it to the policy index.
  5. Detect and remove or quarantine compromised content, then refresh affected indexes.

Explanation

Correct Answers

C:

Source governance helps prevent unreviewed contributions from being promoted to trusted policy evidence.

E:

Correcting source and index state prevents continued retrieval of the known false policy.

Incorrect Answers

A:

Relevance ranking does not authenticate a source or establish policy ownership. A malicious page can match the query well while remaining false.

B:

A weaker tone does not remove the known poisoned source or prevent it from influencing future retrieval.

D:

Copies preserve the same untrusted origin. Counting mirrors does not establish authorized ownership or editorial review of the policy.

Question 398

A saved answer cites a policy URL whose content later changes. What record helps explain what evidence supported the original answer?

  1. A claim that URLs never change their contents.
  2. Only the current page title with no timestamp.
  3. Only the current document revision, overwriting the version recorded when the answer was produced.
  4. The accessed version or content fingerprint, retrieval time, and relevant supporting excerpt under the retention policy.

Explanation

Correct Answer

D:

Version-aware provenance lets reviewers distinguish the evidence originally used from later page revisions. The retention approach should respect access and storage rules while preserving enough evidence for audit. A stable URL alone does not identify the exact content served during the original answer.

Incorrect Answers

A:

Stable addresses can serve updated documents over time.

B:

A title may remain constant across revisions and cannot identify the earlier evidence.

C:

The current revision can differ from the evidence used originally. Replacing the historical reference prevents reconstruction of what supported that answer.

Question 399

A durable note records a temporary promotional discount as if it were permanent. What should the memory design add?

  1. A rule that every saved fact is true forever.
  2. Deletion of all durable notes, including stable project conventions.
  3. The same discount repeated in every future summary.
  4. Validity dates or refresh rules for facts that can expire, with revalidation before use.

Explanation

Correct Answer

D:

Time-bounded state needs lifecycle metadata so a past promotion is not reused as a current entitlement. The note should identify whether the fact remains valid, has expired, or needs refresh. Stable conventions and volatile entitlements should not share an assumption of indefinite truth.

Incorrect Answers

A:

Persistence of storage does not imply persistence of the underlying real-world condition.

B:

Eliminating useful stable memory is unnecessary when volatile facts can be managed explicitly.

C:

Repetition reinforces stale information instead of checking whether it remains applicable.

Question 400

A shared incident summary must exclude personal identifiers while retaining enough information to diagnose a repeated failure across related events. Which THREE practices fit? Choose THREE.

  1. Generate a new pseudonym for the same customer in every event, with no retained mapping for authorized analysis.
  2. Replace all technical failure details with a single generic label, even when the details contain no sensitive data.
  3. Redact identifier columns but publish free-text excerpts without checking them for the same personal information.
  4. Use consistent pseudonymous references where event linkage is needed.
  5. Retain necessary technical error codes and event ordering after reviewing them for sensitive content.
  6. Verify that free-text excerpts do not reintroduce the identifiers removed from structured fields.

Explanation

Correct Answers

D:

Stable nonidentifying labels can preserve relationships without exposing the original personal identifier.

E:

Relevant diagnostic structure can remain while unnecessary personal details are removed.

F:

Sensitive values can survive in narrative text even when dedicated identity fields have been redacted.

Incorrect Answers

A:

This can remove direct identity but also destroys the linkage required to diagnose related events. The scenario calls for consistent references where linkage is necessary.

B:

Removing diagnostic distinctions unnecessarily prevents analysis of recurring causes. Redaction should preserve technical information that is both needed and safe to share.

C:

Free text can repeat names, addresses, or account identifiers removed elsewhere. Field-level redaction alone does not establish that the shared summary excludes personal data.

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