Anthropic CCAR-P (Claude Certified Architect - Professional) Practice Test 24

Question 230

An extractor outputs wrong currency totals although source line items are accurate. Which design choice best matches the stated constraint?

  1. Make retrieval failure explicit and prevent unsupported fallback generation.
  2. Route uncertain complex cases to a stronger model and verify with a difficulty-stratified benchmark.
  3. Delegate arithmetic to a tested calculation component and check units before generation.
  4. Require retrieval-based claim checking and an explicit insufficient-evidence path.

Explanation

Correct Answer

C:

Exact computations should not be trusted to probabilistic text continuation when tools can verify them. Reproduce the failure while checking deterministic numerical correctness at retrieval, tool and model boundaries, changing one hypothesis at a time.

Incorrect Answers

A:

Although grounding-preserving outage behavior can be valuable, it is not the acceptance condition in the question. The missing provision is deterministic numerical correctness, addressed by the control to delegate arithmetic to a tested calculation component and check units before generation.

B:

It is aimed at a different outcome, capability escalation for hard cases. No evidence is given that it provides deterministic numerical correctness; the needed mechanism is to delegate arithmetic to a tested calculation component and check units before generation at the appropriate boundary.

D:

That is primarily a technique for hallucination control on absent evidence. It lacks the explicit measure for deterministic numerical correctness that is necessary here, namely a system that will delegate arithmetic to a tested calculation component and check units before generation.

Question 231

A complex reasoning model fails on a task because the prompt omits the organization's exception policy. Which architecture adjustment is most appropriate?

  1. Test language coverage and revise examples or routing for those locales.
  2. Inspect truncation and context assembly, then prioritize essential fields.
  3. Route uncertain complex cases to a stronger model and verify with a difficulty-stratified benchmark.
  4. Supply the missing authoritative policy in the task context and rerun evaluation.
  5. Audit contradictory demonstrations and add clear boundary cases.

Explanation

Correct Answer

D:

More model capability cannot infer proprietary rules never provided in the input. Reproduce the failure while checking missing domain context diagnosis at retrieval, tool and model boundaries, changing one hypothesis at a time.

Incorrect Answers

A:

The alternative could reduce risk associated with locale-specific failure remediation, yet the described behavior requires missing domain context diagnosis. That outcome depends on the capability to supply the missing authoritative policy in the task context and rerun evaluation.

B:

The proposed measure is aimed at context truncation diagnosis, not missing domain context diagnosis. The architecture must instead supply the missing authoritative policy in the task context and rerun evaluation; a separate change is still needed for missing domain context diagnosis.

C:

The relevant trade-off is not primarily capability escalation for hard cases. A successful fix needs missing domain context diagnosis; otherwise missing domain context diagnosis remains unverified even after this change.

E:

The option belongs to the broader work on repair of contradictory examples, but it cannot by itself deliver missing domain context diagnosis. A decision on this case needs the team to supply the missing authoritative policy in the task context and rerun evaluation and verify that outcome.

Question 232

An agent keeps looping through searches and never finalizes despite finding sufficient evidence. Which approach should be recommended?

  1. Require retrieval-based claim checking and an explicit insufficient-evidence path.
  2. Route uncertain complex cases to a stronger model and verify with a difficulty-stratified benchmark.
  3. Add explicit completion criteria, tool limits and escalation thresholds.
  4. Make retrieval failure explicit and prevent unsupported fallback generation.

Explanation

Correct Answer

C:

An unbounded plan has no reliable stop condition; the remedy is orchestration design, not additional creativity.

Incorrect Answers

A:

The option might improve hallucination control on absent evidence; that remains insufficient for stopping runaway tool loops. The case is resolved by stopping runaway tool loops, which depends on a design that will add explicit completion criteria, tool limits and escalation thresholds.

B:

The proposed action prioritizes capability escalation for hard cases. That can coexist with the right solution, but it cannot replace stopping runaway tool loops: the application must still add explicit completion criteria, tool limits and escalation thresholds.

D:

This addresses grounding-preserving outage behavior; however, the problem asks for stopping runaway tool loops. Without the safeguard to add explicit completion criteria, tool limits and escalation thresholds, the original failure mode is still possible.

Question 233

A concise FAQ prompt works in English but fails on code-mixed regional support messages. What is the best production design response?

  1. Test language coverage and revise examples or routing for those locales.
  2. Supply the missing authoritative policy in the task context and rerun evaluation.
  3. Inspect truncation and context assembly, then prioritize essential fields.
  4. Separate untrusted content, restrict tool access and test prompt-injection cases.

Explanation

Correct Answer

A:

Prompt mismatch may be distribution-specific rather than a global model failure. Reproduce the failure while checking locale-specific failure remediation at retrieval, tool and model boundaries, changing one hypothesis at a time.

Incorrect Answers

B:

Improving missing domain context diagnosis does not imply locale-specific failure remediation. The selected design must test language coverage and revise examples or routing for those locales, whereas this choice omits that direct protection or verification step.

C:

This recommendation targets context truncation diagnosis at a different point in the workflow. It leaves locale-specific failure remediation outside the design contract and does not require the team to test language coverage and revise examples or routing for those locales.

D:

This is a valid architectural concern when optimizing source-content instruction isolation. Here it leaves locale-specific failure remediation unaddressed, whereas the intended change must test language coverage and revise examples or routing for those locales before the case can be considered resolved.

Question 234

A summarizer gives consistently poor answers after a document update, while unchanged prompts previously passed. Which change most directly satisfies this requirement?

  1. Inspect truncation and context assembly, then prioritize essential fields.
  2. Require retrieval-based claim checking and an explicit insufficient-evidence path.
  3. Inspect parsing, index freshness and retrieved versions before retuning the model.
  4. Test language coverage and revise examples or routing for those locales.

Explanation

Correct Answer

C:

An upstream retrieval regression can resemble a model-quality problem in final answers. Reproduce the failure while checking identifying ingestion-induced failures at retrieval, tool and model boundaries, changing one hypothesis at a time.

Incorrect Answers

A:

The control would be justified if context truncation diagnosis were the limiting factor. In this scenario the decisive target is identifying ingestion-induced failures; success requires an architecture that will inspect parsing, index freshness and retrieved versions before retuning the model.

B:

The option belongs to the broader work on hallucination control on absent evidence, but it cannot by itself deliver identifying ingestion-induced failures. A decision on this case needs the team to inspect parsing, index freshness and retrieved versions before retuning the model and verify that outcome.

D:

For a workflow whose primary weakness was locale-specific failure remediation, that proposal could be reasonable. Here the unmet obligation is identifying ingestion-induced failures, and that requires the system to inspect parsing, index freshness and retrieved versions before retuning the model.

Question 235

A model repeatedly treats quoted customer instructions inside retrieved pages as higher-priority commands. Which solution best resolves this operational gap?

  1. Test language coverage and revise examples or routing for those locales.
  2. Supply the missing authoritative policy in the task context and rerun evaluation.
  3. Audit contradictory demonstrations and add clear boundary cases.
  4. Inspect parsing, index freshness and retrieved versions before retuning the model.
  5. Inspect truncation and context assembly, then prioritize essential fields.
  6. Separate untrusted content, restrict tool access and test prompt-injection cases.

Explanation

Correct Answer

F:

The defect concerns trust-boundary enforcement rather than answer wording alone. Reproduce the failure while checking source-content instruction isolation at retrieval, tool and model boundaries, changing one hypothesis at a time.

Incorrect Answers

A:

This addresses locale-specific failure remediation; however, the problem asks for source-content instruction isolation. Without the safeguard to separate untrusted content, restrict tool access and test prompt-injection cases, the original failure mode is still possible.

B:

That choice would be responsive to missing domain context diagnosis, but the reported defect is different. It leaves source-content instruction isolation unproved because there is no commitment to separate untrusted content, restrict tool access and test prompt-injection cases.

C:

There is a legitimate use case for repair of contradictory examples, but it does not supply the test or enforcement needed for source-content instruction isolation. The design needs to separate untrusted content, restrict tool access and test prompt-injection cases.

D:

Consider its actual operating scope: identifying ingestion-induced failures. The requested assurance is source-content instruction isolation, which calls for an implementation that can separate untrusted content, restrict tool access and test prompt-injection cases, not merely for a related improvement.

E:

The alternative could reduce risk associated with context truncation diagnosis, yet the described behavior requires source-content instruction isolation. That outcome depends on the capability to separate untrusted content, restrict tool access and test prompt-injection cases.

Question 236

An answer quality drop appears only when tool responses exceed the input budget. What design decision would best fit these facts?

  1. Inspect parsing, index freshness and retrieved versions before retuning the model.
  2. Supply the missing authoritative policy in the task context and rerun evaluation.
  3. Inspect truncation and context assembly, then prioritize essential fields.
  4. Separate untrusted content, restrict tool access and test prompt-injection cases.

Explanation

Correct Answer

C:

Critical evidence may be lost before inference; increasing prompt specificity cannot restore omitted inputs. Reproduce the failure while checking context truncation diagnosis at retrieval, tool and model boundaries, changing one hypothesis at a time.

Incorrect Answers

A:

This is a valid architectural concern when optimizing identifying ingestion-induced failures. Here it leaves context truncation diagnosis unaddressed, whereas the intended change must inspect truncation and context assembly, then prioritize essential fields before the case can be considered resolved.

B:

This would make sense as an additional project activity for missing domain context diagnosis. The priority required here is context truncation diagnosis, so the workflow still needs a control that will inspect truncation and context assembly, then prioritize essential fields.

D:

Its main effect concerns source-content instruction isolation. The case explicitly hinges on context truncation diagnosis; to resolve it reliably, the implementation has to inspect truncation and context assembly, then prioritize essential fields, which this choice does not establish.

Question 237

A classifier is inconsistent because its template contains conflicting examples for the same labels. A separate prompt template inserts user text without escaping its example delimiters, so data can be mistaken for a new instruction or demonstration. Which TWO diagnoses lead to direct fixes? Choose TWO.

  1. Reproduce delimiter collisions and enforce escaping or structured boundaries between template instructions and inserted data.
  2. Audit contradictory demonstrations and add clear boundary cases.
  3. Trust the model to ignore inserted delimiter text and add a refusal reminder without fixing the template escaping.
  4. Test language coverage and revise examples or routing for those locales.
  5. Delegate arithmetic to a tested calculation component and check units before generation.

Explanation

Correct Answers

A:

An unescaped data field can change how the prompt is parsed conceptually, even when the intended examples are correct. Robust delimiting keeps user content in the data role and prevents accidental injection of new examples.

B:

Contradictory labeled demonstrations give the model incompatible signals about the decision boundary. The template owner should resolve conflicts, add edge cases and version the corrected examples before concluding that the model itself is unsuitable.

Incorrect Answers

C:

A prompt reminder does not repair ambiguous boundaries between untrusted data and examples. Input escaping or structured isolation must be enforced before the model sees the assembled prompt.

D:

It emphasizes locale-specific failure remediation without closing the stated gap in repair of contradictory examples and safe templating boundaries for inserted user data. By contrast, the remedy to reproduce delimiter collisions and enforce escaping or structured boundaries between template instructions and inserted data acts on the condition that makes this scenario fail.

E:

That is primarily a technique for deterministic numerical correctness. It lacks the explicit measure for repair of contradictory examples and safe templating boundaries for inserted user data that is necessary here, namely a system that will audit contradictory demonstrations and add clear boundary cases.

Question 238

A simpler model starts misclassifying edge cases after a cost optimization. What should the solution architect prioritize?

  1. Supply the missing authoritative policy in the task context and rerun evaluation.
  2. Delegate arithmetic to a tested calculation component and check units before generation.
  3. Require retrieval-based claim checking and an explicit insufficient-evidence path.
  4. Route uncertain complex cases to a stronger model and verify with a difficulty-stratified benchmark.
  5. Add explicit completion criteria, tool limits and escalation thresholds.

Explanation

Correct Answer

D:

Model mismatch is best diagnosed by task capability needs, not by assuming all failures are random. Reproduce the failure while checking capability escalation for hard cases at retrieval, tool and model boundaries, changing one hypothesis at a time.

Incorrect Answers

A:

Consider its actual operating scope: missing domain context diagnosis. The requested assurance is capability escalation for hard cases, which calls for an implementation that can route uncertain complex cases to a stronger model and verify with a difficulty-stratified benchmark, not merely for a related improvement.

B:

The alternative could reduce risk associated with deterministic numerical correctness, yet the described behavior requires capability escalation for hard cases. That outcome depends on the capability to route uncertain complex cases to a stronger model and verify with a difficulty-stratified benchmark.

C:

The proposed measure is aimed at hallucination control on absent evidence, not capability escalation for hard cases. The architecture must instead route uncertain complex cases to a stronger model and verify with a difficulty-stratified benchmark; a separate change is still needed for capability escalation for hard cases.

E:

The control would be justified if stopping runaway tool loops were the limiting factor. In this scenario the decisive target is capability escalation for hard cases; success requires an architecture that will route uncertain complex cases to a stronger model and verify with a difficulty-stratified benchmark.

Question 239

A refund-policy assistant quotes a scanned customer form as allowing a refund, but the underlying clause says “not eligible for refunds.” Investigation shows that optical character recognition sometimes drops short negation words. What remediation best targets this failure?

  1. Increase the number of retrieval results without verifying the extracted text of the disputed clause.
  2. Add extraction checks for polarity-changing words and regression cases that fail when a scanned policy’s negation is lost.
  3. Tune the model for more enthusiastic customer-service phrasing to compensate for uncertain source text.
  4. Cache the incorrectly extracted policy passage so all users receive a consistent answer.

Explanation

Correct Answer

B:

The incorrect answer originates before generation: extraction changes the legal meaning of the evidence. Negation-sensitive OCR tests, selective visual verification and versioned fixtures reduce unsupported eligibility conclusions.

Incorrect Answers

A:

More chunks cannot restore a missing negation in a corrupt source passage. The extraction step must preserve the language that determines eligibility.

C:

Tone adjustment does not correct polarity errors and risks a more persuasive false promise. The evidence-processing defect needs direct tests and an abstention path for unreadable clauses.

D:

Consistency is not correctness. Reusing a passage that omits a controlling negation propagates the same compliance error more broadly.

Question 240

A service passes the full 80-kilobyte diagnostic payload from each tool call back into its Claude context, although the final answer requires only three validated fields. Token costs rise with every call. What is the most targeted optimization?

  1. Normalize tool responses into the required typed fields and preserve an auditable reference to full logs outside the model context.
  2. Raise the maximum output-token setting to allow the model to condense the entire 80-kilobyte payload.
  3. Switch to a cheaper model without changing the payload passed into context.
  4. Place the unfiltered diagnostic logs in the system instruction so they remain visible throughout the session.

Explanation

Correct Answer

A:

A gateway-side projection strips irrelevant diagnostic bytes before they become input tokens while preserving the evidence path for authorized audit. Test the projection against cases where omitted fields might affect correctness.

Incorrect Answers

B:

Increasing output tokens does not shrink repeated input payloads, and model-side summarization still requires paying for every source byte presented. Filter results before context assembly instead.

C:

A less expensive model can change unit economics but cannot remove avoidable context growth or prevent truncation. Reducing unnecessary tool-result tokens addresses the identified cost driver directly.

D:

System instructions still consume context and should not become a dump of transient diagnostics. Persistent raw logs also increase accidental disclosure risks without improving the three-field response contract.

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