Anthropic CCAR-P (Claude Certified Architect - Professional) Practice Test 27

Question 260

An agent occasionally triggers duplicate writes after timeout but no metric tracks repeated transaction keys. Which change most directly satisfies this requirement?

  1. Monitor source-to-index lag and stale-citation rates.
  2. Tag traces with prompt, model, retrieval and tool versions for rollback comparison.
  3. Track p95 and p99 end-to-end latency along with timeout rates.
  4. Alert on deduplicated retries and post-commit replay attempts.

Explanation

Correct Answer

D:

Side-effect telemetry catches operational failures that response text metrics cannot see. Segment the signal for production side-effect integrity monitoring by release, workload and affected population, with alerts tied to an assigned responder.

Incorrect Answers

A:

This is a valid architectural concern when optimizing continuous source freshness assurance. Here it leaves production side-effect integrity monitoring unaddressed, whereas the intended change must alert on deduplicated retries and post-commit replay attempts before the case can be considered resolved.

B:

This would make sense as an additional project activity for version-to-quality correlation. The priority required here is production side-effect integrity monitoring, so the workflow still needs a control that will alert on deduplicated retries and post-commit replay attempts.

C:

By itself, this improves tail performance monitoring, but the case is not asking the architect to optimize that dimension. The requirement is production side-effect integrity monitoring, supported by an approach to alert on deduplicated retries and post-commit replay attempts.

Question 261

An HR assistant is forbidden from exposing salary records but a policy prompt alone controls access. Which solution best resolves this operational gap?

  1. Minimize sensitive inputs and apply output checks for protected fields.
  2. Move sensitivity classification and entitlement checks before context injection.
  3. Validate against an allowed output schema and reject or repair failures.
  4. Enforce document and tool authorization at the retrieval boundary before generating answers.

Explanation

Correct Answer

D:

A guardrail that merely requests discretion cannot prevent the model from receiving unauthorized data. Test access enforcement ahead of inference with adversarial cases and an application-enforced boundary; credentials and service policies still have to enforce access enforcement ahead of inference.

Incorrect Answers

A:

For a workflow whose primary weakness was multiple layers of sensitive-data protection, that proposal could be reasonable. Here the unmet obligation is access enforcement ahead of inference, and that requires the system to enforce document and tool authorization at the retrieval boundary before generating answers.

B:

Although preventive filtering before model context can be valuable, it is not the acceptance condition in the question. The missing provision is access enforcement ahead of inference, addressed by the control to enforce document and tool authorization at the retrieval boundary before generating answers.

C:

It is aimed at a different outcome, application-layer output enforcement. No evidence is given that it provides access enforcement ahead of inference; the needed mechanism is to enforce document and tool authorization at the retrieval boundary before generating answers at the appropriate boundary.

Question 262

A customer-service agent can issue refunds and should not exceed an employee's authorization limit. What design decision would best fit these facts?

  1. Mark retrieved text as untrusted data and isolate executable tool privileges.
  2. Check refund amount, caller scope and approvals in a backend policy gateway.
  3. Require explicit authenticated approval before irreversible financial commits.
  4. Constrain network destinations, tool arguments and egress privileges to approved endpoints.

Explanation

Correct Answer

B:

Transaction controls belong in the system of record; natural-language instructions are not sufficient authorization.

Incorrect Answers

A:

Consider its actual operating scope: untrusted-content isolation. The requested assurance is server-side limits on financial actions, which calls for an implementation that can check refund amount, caller scope and approvals in a backend policy gateway, not merely for a related improvement.

C:

The proposed measure is aimed at transaction human-approval boundary, not server-side limits on financial actions. The architecture must instead check refund amount, caller scope and approvals in a backend policy gateway; a separate change is still needed for server-side limits on financial actions.

D:

The relevant trade-off is not primarily approved network egress boundaries. A successful fix needs server-side limits on financial actions; otherwise server-side limits on financial actions remains unverified even after this change.

Question 263

A medical information assistant must escalate unsupported diagnostic conclusions instead of guessing. Its output also must never contradict an externally maintained list of dangerous drug contraindications. Which TWO layered safeguards are needed? Choose TWO.

  1. Require explicit authenticated approval before irreversible financial commits.
  2. Rely on a system-prompt reminder about unsafe diagnoses and contraindications instead of an external clinical rule check.
  3. Calibrate safety rules by use case and evaluate both harmful completions and false refusals.
  4. Enforce contraindication rules deterministically against a trusted, maintained clinical rule source before displaying the response.
  5. Ground approved information, restrict clinical conclusions and route high-risk cases to clinicians.

Explanation

Correct Answers

D:

High-severity exclusions should be checked by a separately governed rule service, not left to a generative instruction. Its rule version and decision can be audited even when the model otherwise produces a fluent answer.

E:

Unsupported diagnostic advice requires evidence constraints and a clinical escalation path rather than a confident generated answer. Separately, an external deterministic contraindication rule makes a high-severity exclusion enforceable even when the model sounds plausible.

Incorrect Answers

A:

Its main effect concerns transaction human-approval boundary. The case explicitly hinges on high-stakes content escalation and independent deterministic contraindication enforcement; to resolve it reliably, the implementation has to ground approved information, restrict clinical conclusions and route high-risk cases to clinicians, which this choice does not establish.

B:

Prompts can shape the assistant’s behavior but cannot enforce a continuously maintained contraindication table. The requirement demands a separately validated rules decision and qualified escalation for unsupported claims.

C:

The proposed action prioritizes context-specific safety tuning. That can coexist with the right solution, but it cannot replace high-stakes content escalation and independent deterministic contraindication enforcement: the application must still ground approved information, restrict clinical conclusions and route high-risk cases to clinicians.

Question 264

A publicly accessible assistant can be prompted to output personal data found in uploaded context. What should the solution architect prioritize?

  1. Minimize sensitive inputs and apply output checks for protected fields.
  2. Add citation checks, calibrated uncertainty and escalation guidance to the user interface.
  3. Enforce document and tool authorization at the retrieval boundary before generating answers.
  4. Validate against an allowed output schema and reject or repair failures.

Explanation

Correct Answer

A:

Input access control is primary, while output validation is a second line of defense against leakage. Test multiple layers of sensitive-data protection with adversarial cases and an application-enforced boundary; credentials and service policies still have to enforce multiple layers of sensitive-data protection.

Incorrect Answers

B:

That is primarily a technique for safe communication of model uncertainty. It lacks the explicit measure for multiple layers of sensitive-data protection that is necessary here, namely a system that will minimize sensitive inputs and apply output checks for protected fields.

C:

Improving access enforcement ahead of inference does not imply multiple layers of sensitive-data protection. The selected design must minimize sensitive inputs and apply output checks for protected fields, whereas this choice omits that direct protection or verification step.

D:

This recommendation targets application-layer output enforcement at a different point in the workflow. It leaves multiple layers of sensitive-data protection outside the design contract and does not require the team to minimize sensitive inputs and apply output checks for protected fields.

Question 265

A service ingests hostile third-party documents containing instructions to disable monitoring. Which design choice best matches the stated constraint?

  1. Enforce document and tool authorization at the retrieval boundary before generating answers.
  2. Constrain network destinations, tool arguments and egress privileges to approved endpoints.
  3. Mark retrieved text as untrusted data and isolate executable tool privileges.
  4. Move sensitivity classification and entitlement checks before context injection.

Explanation

Correct Answer

C:

Prompt injection crosses a trust boundary; tool-gateway permission checks contain its consequences. Test untrusted-content isolation with adversarial cases and an application-enforced boundary; credentials and service policies still have to enforce untrusted-content isolation.

Incorrect Answers

A:

The relevant trade-off is not primarily access enforcement ahead of inference. A successful fix needs untrusted-content isolation; otherwise untrusted-content isolation remains unverified even after this change.

B:

The control would be justified if approved network egress boundaries were the limiting factor. In this scenario the decisive target is untrusted-content isolation; success requires an architecture that will mark retrieved text as untrusted data and isolate executable tool privileges.

D:

This shifts attention toward preventive filtering before model context when the relevant decision is about untrusted-content isolation. Unless the architecture can mark retrieved text as untrusted data and isolate executable tool privileges, the failure described in the stem remains unresolved.

Question 266

A procurement agent executes a critical purchase immediately after classifying it as low risk. Which architecture adjustment is most appropriate?

  1. Check refund amount, caller scope and approvals in a backend policy gateway.
  2. Require explicit authenticated approval before irreversible financial commits.
  3. Enforce document and tool authorization at the retrieval boundary before generating answers.
  4. Constrain network destinations, tool arguments and egress privileges to approved endpoints.
  5. Ground approved information, restrict clinical conclusions and route high-risk cases to clinicians.

Explanation

Correct Answer

B:

Risk labels produced by a model do not confer legal or financial authority. Test transaction human-approval boundary with adversarial cases and an application-enforced boundary; credentials and service policies still have to enforce transaction human-approval boundary.

Incorrect Answers

A:

The design concern is real but different: server-side limits on financial actions. It does not impose the specific requirement for transaction human-approval boundary; a suitable solution must require explicit authenticated approval before irreversible financial commits and demonstrate the result.

C:

That choice would be responsive to access enforcement ahead of inference, but the reported defect is different. It leaves transaction human-approval boundary unproved because there is no commitment to require explicit authenticated approval before irreversible financial commits.

D:

There is a legitimate use case for approved network egress boundaries, but it does not supply the test or enforcement needed for transaction human-approval boundary. The design needs to require explicit authenticated approval before irreversible financial commits.

E:

Consider its actual operating scope: high-stakes content escalation. The requested assurance is transaction human-approval boundary, which calls for an implementation that can require explicit authenticated approval before irreversible financial commits, not merely for a related improvement.

Question 267

A team applies one aggressive refusal policy to every product, blocking benign internal summaries. Which approach should be recommended?

  1. Add citation checks, calibrated uncertainty and escalation guidance to the user interface.
  2. Enforce document and tool authorization at the retrieval boundary before generating answers.
  3. Check refund amount, caller scope and approvals in a backend policy gateway.
  4. Calibrate safety rules by use case and evaluate both harmful completions and false refusals.

Explanation

Correct Answer

D:

Risk controls should target the actual threat and be measured for collateral utility loss. Test context-specific safety tuning with adversarial cases and an application-enforced boundary; credentials and service policies still have to enforce context-specific safety tuning.

Incorrect Answers

A:

This recommendation targets safe communication of model uncertainty at a different point in the workflow. It leaves context-specific safety tuning outside the design contract and does not require the team to calibrate safety rules by use case and evaluate both harmful completions and false refusals.

B:

This is a valid architectural concern when optimizing access enforcement ahead of inference. Here it leaves context-specific safety tuning unaddressed, whereas the intended change must calibrate safety rules by use case and evaluate both harmful completions and false refusals before the case can be considered resolved.

C:

This would make sense as an additional project activity for server-side limits on financial actions. The priority required here is context-specific safety tuning, so the workflow still needs a control that will calibrate safety rules by use case and evaluate both harmful completions and false refusals.

Question 268

An assistant produces JSON that passes a prompt instruction but occasionally contains unauthorized field names. What is the best production design response?

  1. Enforce document and tool authorization at the retrieval boundary before generating answers.
  2. Enforce contraindication rules deterministically against a trusted, maintained clinical rule source before displaying the response.
  3. Minimize sensitive inputs and apply output checks for protected fields.
  4. Validate against an allowed output schema and reject or repair failures.

Explanation

Correct Answer

D:

Structured contract enforcement must occur outside model generation for dependable downstream use. Test application-layer output enforcement with adversarial cases and an application-enforced boundary; credentials and service policies still have to enforce application-layer output enforcement.

Incorrect Answers

A:

This shifts attention toward access enforcement ahead of inference when the relevant decision is about application-layer output enforcement. Unless the architecture can validate against an allowed output schema and reject or repair failures, the failure described in the stem remains unresolved.

B:

For a workflow whose primary weakness was independent deterministic contraindication enforcement, that proposal could be reasonable. Here the unmet obligation is application-layer output enforcement, and that requires the system to validate against an allowed output schema and reject or repair failures.

C:

Although multiple layers of sensitive-data protection can be valuable, it is not the acceptance condition in the question. The missing provision is application-layer output enforcement, addressed by the control to validate against an allowed output schema and reject or repair failures.

Question 269

A service displays AI answers as authoritative policy without references or uncertainty indicators. Which change most directly satisfies this requirement?

  1. Check refund amount, caller scope and approvals in a backend policy gateway.
  2. Mark retrieved text as untrusted data and isolate executable tool privileges.
  3. Add citation checks, calibrated uncertainty and escalation guidance to the user interface.
  4. Enforce document and tool authorization at the retrieval boundary before generating answers.

Explanation

Correct Answer

C:

Presentation design is a safety control when users may over-trust plausible language. Test safe communication of model uncertainty with adversarial cases and an application-enforced boundary; credentials and service policies still have to enforce safe communication of model uncertainty.

Incorrect Answers

A:

There is a legitimate use case for server-side limits on financial actions, but it does not supply the test or enforcement needed for safe communication of model uncertainty. The design needs to add citation checks, calibrated uncertainty and escalation guidance to the user interface.

B:

This addresses untrusted-content isolation; however, the problem asks for safe communication of model uncertainty. Without the safeguard to add citation checks, calibrated uncertainty and escalation guidance to the user interface, the original failure mode is still possible.

D:

The proposed measure is aimed at access enforcement ahead of inference, not safe communication of model uncertainty. The architecture must instead add citation checks, calibrated uncertainty and escalation guidance to the user interface; a separate change is still needed for safe communication of model uncertainty.

Question 270

An agent can reach arbitrary external URLs through a general browsing connector. Which solution best resolves this operational gap?

  1. Constrain network destinations, tool arguments and egress privileges to approved endpoints.
  2. Mark retrieved text as untrusted data and isolate executable tool privileges.
  3. Enforce document and tool authorization at the retrieval boundary before generating answers.
  4. Check refund amount, caller scope and approvals in a backend policy gateway.

Explanation

Correct Answer

A:

Network-level restrictions limit data exfiltration routes independent of the model's intent. Test approved network egress boundaries with adversarial cases and an application-enforced boundary; credentials and service policies still have to enforce approved network egress boundaries.

Incorrect Answers

B:

Its main effect concerns untrusted-content isolation. The case explicitly hinges on approved network egress boundaries; to resolve it reliably, the implementation has to constrain network destinations, tool arguments and egress privileges to approved endpoints, which this choice does not establish.

C:

The option might improve access enforcement ahead of inference; that remains insufficient for approved network egress boundaries. The case is resolved by approved network egress boundaries, which depends on a design that will constrain network destinations, tool arguments and egress privileges to approved endpoints.

D:

The proposed action prioritizes server-side limits on financial actions. That can coexist with the right solution, but it cannot replace approved network egress boundaries: the application must still constrain network destinations, tool arguments and egress privileges to approved endpoints.

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