Anthropic CCAR-P (Claude Certified Architect - Professional) Practice Test 29

Question 280

A regulatory team accepts a model's cited answer without checking whether cited pages support the claims. What design decision would best fit these facts?

  1. Identify stale-evidence risk and add validity-date controls to retrieval.
  2. Treat instruction laundering through retrieved content as a distinct adversarial threat.
  3. Model mid-execution authorization drift and require rechecks before sensitive commits.
  4. Include citation-faithfulness failures in the hazard analysis and test plan.
  5. Monitor both under-refusal harms and false refusals after guardrail changes.

Explanation

Correct Answer

D:

A reference marker may look reassuring while pointing to irrelevant evidence. Assess threats to false grounding and citation laundering risk by consequence, detectability and exploit path; an apparently successful answer can still hide failure to deliver false grounding and citation laundering risk.

Incorrect Answers

A:

The option might improve exposure to outdated authoritative sources; that remains insufficient for false grounding and citation laundering risk. The case is resolved by false grounding and citation laundering risk, which depends on a design that will include citation-faithfulness failures in the hazard analysis and test plan.

B:

The proposed action prioritizes prompt-injection trust-boundary exposure. That can coexist with the right solution, but it cannot replace false grounding and citation laundering risk: the application must still include citation-faithfulness failures in the hazard analysis and test plan.

C:

The design concern is real but different: revoked authority during autonomous execution. It does not impose the specific requirement for false grounding and citation laundering risk; a suitable solution must include citation-faithfulness failures in the hazard analysis and test plan and demonstrate the result.

E:

That choice would be responsive to safety regression hidden by usability gains, but the reported defect is different. It leaves false grounding and citation laundering risk unproved because there is no commitment to include citation-faithfulness failures in the hazard analysis and test plan.

Question 281

An insurer automates rejection notices and has no channel for customers to contest errors. What should the solution architect prioritize?

  1. Treat unsupported numerical interpolation as a high-severity hallucination requiring deterministic verification.
  2. Identify accountability and appeal risk before production release.
  3. Audit data, outcomes and subgroup disparities before relying on automated ranking.
  4. Identify stale-evidence risk and add validity-date controls to retrieval.

Explanation

Correct Answer

B:

High-impact outcomes require routes for explanation, correction and accountable review. Assess threats to missing appeal and remediation process by consequence, detectability and exploit path; an apparently successful answer can still hide failure to deliver missing appeal and remediation process.

Incorrect Answers

A:

That is primarily a technique for fabricated quantitative evidence risk. It lacks the explicit measure for missing appeal and remediation process that is necessary here, namely a system that will identify accountability and appeal risk before production release.

C:

This recommendation targets bias propagation from training labels at a different point in the workflow. It leaves missing appeal and remediation process outside the design contract and does not require the team to identify accountability and appeal risk before production release.

D:

This is a valid architectural concern when optimizing exposure to outdated authoritative sources. Here it leaves missing appeal and remediation process unaddressed, whereas the intended change must identify accountability and appeal risk before production release before the case can be considered resolved.

Question 282

A model's refusal rate drops after a system change, but sensitive content incidents increase. Which design choice best matches the stated constraint?

  1. Evaluate operational and logging paths as part of the attack surface.
  2. Model mid-execution authorization drift and require rechecks before sensitive commits.
  3. Prioritize cross-tenant action risk using severity and privilege boundaries, not raw frequency.
  4. Include citation-faithfulness failures in the hazard analysis and test plan.
  5. Monitor both under-refusal harms and false refusals after guardrail changes.
  6. Count failed and abandoned work in reliability and risk reporting.

Explanation

Correct Answer

E:

A single apparent utility gain can mask an expanded safety failure surface. Assess threats to safety regression hidden by usability gains by consequence, detectability and exploit path; an apparently successful answer can still hide failure to deliver safety regression hidden by usability gains.

Incorrect Answers

A:

The control would be justified if non-model information disclosure paths were the limiting factor. In this scenario the decisive target is safety regression hidden by usability gains; success requires an architecture that will monitor both under-refusal harms and false refusals after guardrail changes.

B:

The option belongs to the broader work on revoked authority during autonomous execution, but it cannot by itself deliver safety regression hidden by usability gains. A decision on this case needs the team to monitor both under-refusal harms and false refusals after guardrail changes and verify that outcome.

C:

This shifts attention toward severity-sensitive tenant isolation risk when the relevant decision is about safety regression hidden by usability gains. Unless the architecture can monitor both under-refusal harms and false refusals after guardrail changes, the failure described in the stem remains unresolved.

D:

For a workflow whose primary weakness was false grounding and citation laundering risk, that proposal could be reasonable. Here the unmet obligation is safety regression hidden by usability gains, and that requires the system to monitor both under-refusal harms and false refusals after guardrail changes.

F:

It is aimed at a different outcome, honest measurement of silent timeouts. No evidence is given that it provides safety regression hidden by usability gains; the needed mechanism is to monitor both under-refusal harms and false refusals after guardrail changes at the appropriate boundary.

Question 283

A company stores raw prompts indefinitely to aid debugging despite frequent customer secrets. Which architecture adjustment is most appropriate?

  1. Include retention and secondary-use exposure in its risk assessment.
  2. Treat unsupported numerical interpolation as a high-severity hallucination requiring deterministic verification.
  3. Prioritize cross-tenant action risk using severity and privilege boundaries, not raw frequency.
  4. Evaluate operational and logging paths as part of the attack surface.

Explanation

Correct Answer

A:

Diagnostics can create a long-lived sensitive-data asset separate from the main inference path. Assess threats to operational data retention risk by consequence, detectability and exploit path; an apparently successful answer can still hide failure to deliver operational data retention risk.

Incorrect Answers

B:

That choice would be responsive to fabricated quantitative evidence risk, but the reported defect is different. It leaves operational data retention risk unproved because there is no commitment to include retention and secondary-use exposure in its risk assessment.

C:

There is a legitimate use case for severity-sensitive tenant isolation risk, but it does not supply the test or enforcement needed for operational data retention risk. The design needs to include retention and secondary-use exposure in its risk assessment.

D:

Consider its actual operating scope: non-model information disclosure paths. The requested assurance is operational data retention risk, which calls for an implementation that can include retention and secondary-use exposure in its risk assessment, not merely for a related improvement.

Question 284

An assistant can summarize claim files safely, but final claim denial has legal consequences. Which next step is most defensible?

  1. Distinguish sampled quality assurance from per-decision approval in release records.
  2. Provide evidence-linked diffs, highlighted uncertainties and explicit approve/reject actions.
  3. Require qualified human approval before sending an adverse decision.
  4. Keep dual authorization external to the model and store approval identities in an audit trail.

Explanation

Correct Answer

C:

The irreversible high-impact action needs accountable review even if upstream analysis is automated. Record who authorized human decision ownership for adverse actions and keep rejection and timeout state durable so retries cannot bypass an approval decision.

Incorrect Answers

A:

This is a valid architectural concern when optimizing truthful evidence of human review scope. Here it leaves human decision ownership for adverse actions unaddressed, whereas the intended change must require qualified human approval before sending an adverse decision before the case can be considered resolved.

B:

This would make sense as an additional project activity for informed review with evidence. The priority required here is human decision ownership for adverse actions, so the workflow still needs a control that will require qualified human approval before sending an adverse decision.

D:

Its main effect concerns audited multi-party access approval. The case explicitly hinges on human decision ownership for adverse actions; to resolve it reliably, the implementation has to require qualified human approval before sending an adverse decision, which this choice does not establish.

Question 285

A finance workflow requires review for amounts above a fixed threshold while small expenses may auto-process. Which approach should be recommended?

  1. Fail closed for mandatory approvals and notify an authorized fallback reviewer.
  2. Redesign the review workload and require meaningful evidence inspection for critical cases.
  3. Enforce a deterministic amount-based approval gate in the workflow engine.
  4. Route uncertain or high-risk cases to clinicians using calibrated thresholds and workload monitoring.

Explanation

Correct Answer

C:

Threshold routing must be reproducible and cannot rely solely on a model's risk self-assessment. Record who authorized rule-governed risk-based approval and keep rejection and timeout state durable so retries cannot bypass an approval decision.

Incorrect Answers

A:

For a workflow whose primary weakness was approval timeout safety, that proposal could be reasonable. Here the unmet obligation is rule-governed risk-based approval, and that requires the system to enforce a deterministic amount-based approval gate in the workflow engine.

B:

Although avoidance of review automation bias can be valuable, it is not the acceptance condition in the question. The missing provision is rule-governed risk-based approval, addressed by the control to enforce a deterministic amount-based approval gate in the workflow engine.

D:

It emphasizes confidence- and severity-based review allocation without closing the stated gap in rule-governed risk-based approval. By contrast, the remedy to enforce a deterministic amount-based approval gate in the workflow engine acts on the condition that makes this scenario fail.

Question 286

A physician reviewer gets hundreds of easy cases and misses the truly ambiguous model outputs. What is the best production design response?

  1. Require qualified human approval before sending an adverse decision.
  2. Route uncertain or high-risk cases to clinicians using calibrated thresholds and workload monitoring.
  3. Enforce a deterministic amount-based approval gate in the workflow engine.
  4. Redesign the review workload and require meaningful evidence inspection for critical cases.

Explanation

Correct Answer

B:

Selective human review focuses scarce expertise on cases with the largest expected risk reduction. Record who authorized confidence- and severity-based review allocation and keep rejection and timeout state durable so retries cannot bypass an approval decision.

Incorrect Answers

A:

Consider its actual operating scope: human decision ownership for adverse actions. The requested assurance is confidence- and severity-based review allocation, which calls for an implementation that can route uncertain or high-risk cases to clinicians using calibrated thresholds and workload monitoring, not merely for a related improvement.

C:

The proposed measure is aimed at rule-governed risk-based approval, not confidence- and severity-based review allocation. The architecture must instead route uncertain or high-risk cases to clinicians using calibrated thresholds and workload monitoring; a separate change is still needed for confidence- and severity-based review allocation.

D:

The relevant trade-off is not primarily avoidance of review automation bias. A successful fix needs confidence- and severity-based review allocation; otherwise confidence- and severity-based review allocation remains unverified even after this change.

Question 287

A regulated filing system lets reviewers approve AI drafts without seeing the source evidence. It also lets the document requester approve their own filing under a shared reviewer account. Which TWO changes create a meaningful review gate? Choose TWO.

  1. Distinguish sampled quality assurance from per-decision approval in release records.
  2. Enforce separation of requester and approver identities in the workflow engine and disallow self-approval.
  3. Require qualified human approval before sending an adverse decision.
  4. Authenticate users as reviewers but allow a requester to switch into the same reviewer role for their own submission.
  5. Provide evidence-linked diffs, highlighted uncertainties and explicit approve/reject actions.

Explanation

Correct Answers

B:

Approval is not independent when the same principal can originate and authorize a filing. The service must verify distinct authenticated actors and store the authorization trail rather than relying on the model to infer the reviewer role.

E:

Approval is only meaningful when a reviewer can inspect the proposed changes against evidence and explicitly accept or reject them. Showing source-linked diffs and uncertainties supports that judgment; identity separation must also prevent the requester from granting approval to themselves.

Incorrect Answers

A:

Its main effect concerns truthful evidence of human review scope. The case explicitly hinges on informed review with evidence and independent actor enforcement for approvals; to resolve it reliably, the implementation has to provide evidence-linked diffs, highlighted uncertainties and explicit approve/reject actions, which this choice does not establish.

C:

The proposed action prioritizes human decision ownership for adverse actions. That can coexist with the right solution, but it cannot replace informed review with evidence and independent actor enforcement for approvals: the application must still provide evidence-linked diffs, highlighted uncertainties and explicit approve/reject actions.

D:

Verifying that a requester also holds a reviewer role is not separation of duties. The service must prohibit approval of one’s own case and keep distinct authenticated approval identities in durable records.

Question 288

An agent automatically retries a destructive operation after a human rejected it. What design decision would best fit these facts?

  1. Require qualified human approval before sending an adverse decision.
  2. Separate recommendation, approval and execution with authenticated actor identities.
  3. Persist rejection state and prevent automatic re-entry without new authorization.
  4. Distinguish sampled quality assurance from per-decision approval in release records.

Explanation

Correct Answer

C:

Approval is a stateful business decision; retries must not circumvent a denial. Record who authorized durable human veto enforcement and keep rejection and timeout state durable so retries cannot bypass an approval decision.

Incorrect Answers

A:

It emphasizes human decision ownership for adverse actions without closing the stated gap in durable human veto enforcement. By contrast, the remedy to persist rejection state and prevent automatic re-entry without new authorization acts on the condition that makes this scenario fail.

B:

That is primarily a technique for segregation between drafting and committing. It lacks the explicit measure for durable human veto enforcement that is necessary here, namely a system that will persist rejection state and prevent automatic re-entry without new authorization.

D:

This recommendation targets truthful evidence of human review scope at a different point in the workflow. It leaves durable human veto enforcement outside the design contract and does not require the team to persist rejection state and prevent automatic re-entry without new authorization.

Question 289

A quality team labels all model responses 'human reviewed' after a random 1% audit. Which intervention most directly addresses the failure?

  1. Require qualified human approval before sending an adverse decision.
  2. Distinguish sampled quality assurance from per-decision approval in release records.
  3. Redesign the review workload and require meaningful evidence inspection for critical cases.
  4. Provide evidence-linked diffs, highlighted uncertainties and explicit approve/reject actions.

Explanation

Correct Answer

B:

Statistical sampling estimates system quality but does not authorize individual irreversible actions. Record who authorized truthful evidence of human review scope and keep rejection and timeout state durable so retries cannot bypass an approval decision.

Incorrect Answers

A:

The relevant trade-off is not primarily human decision ownership for adverse actions. A successful fix needs truthful evidence of human review scope; otherwise truthful evidence of human review scope remains unverified even after this change.

C:

The option belongs to the broader work on avoidance of review automation bias, but it cannot by itself deliver truthful evidence of human review scope. A decision on this case needs the team to distinguish sampled quality assurance from per-decision approval in release records and verify that outcome.

D:

This shifts attention toward informed review with evidence when the relevant decision is about truthful evidence of human review scope. Unless the architecture can distinguish sampled quality assurance from per-decision approval in release records, the failure described in the stem remains unresolved.

Question 290

A reviewer repeatedly rubber-stamps model-generated high-risk summaries under a strict 10-second quota. What should the solution architect prioritize?

  1. Redesign the review workload and require meaningful evidence inspection for critical cases.
  2. Require qualified human approval before sending an adverse decision.
  3. Provide evidence-linked diffs, highlighted uncertainties and explicit approve/reject actions.
  4. Route uncertain or high-risk cases to clinicians using calibrated thresholds and workload monitoring.

Explanation

Correct Answer

A:

Nominal human presence is not an effective control if the process prevents independent judgment. Record who authorized avoidance of review automation bias and keep rejection and timeout state durable so retries cannot bypass an approval decision.

Incorrect Answers

B:

This addresses human decision ownership for adverse actions; however, the problem asks for avoidance of review automation bias. Without the safeguard to redesign the review workload and require meaningful evidence inspection for critical cases, the original failure mode is still possible.

C:

That choice would be responsive to informed review with evidence, but the reported defect is different. It leaves avoidance of review automation bias unproved because there is no commitment to redesign the review workload and require meaningful evidence inspection for critical cases.

D:

There is a legitimate use case for confidence- and severity-based review allocation, but it does not supply the test or enforcement needed for avoidance of review automation bias. The design needs to redesign the review workload and require meaningful evidence inspection for critical cases.

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