Anthropic CCAR-P (Claude Certified Architect - Professional) Practice Test 35

Question 340

The energy operations release team needs a numerical gate for Data exposure and retention review. Rather than equating deployment with success, it wants to catch cases in which encryption in transit alone does not resolve overcollection or retention. Which indicator should it track?

  1. Measure reviewer disagreement and caught-error rates in seeded tests
  2. Measure successful cross-boundary actions on red-team scenarios
  3. Measure policy regression detection time after controlled changes
  4. Measure calibrated confidence and source-supported assertions
  5. Measure unauthorized data persistence and access findings

Explanation

Correct Answer

E:

The correct signal is unauthorized data persistence and access findings; it must be computed for representative workloads and compared with a baseline. It checks whether the control layer can trace data paths, access, retention and deletion obligations before launch while surfacing the exposure that encryption in transit alone does not resolve overcollection or retention.

Incorrect Answers

A:

When reviewers rubber-stamp model recommendations under time pressure, the control layer can design independent checks and reason-coded overrides for high-stakes cases; but the acceptance gate must instead inspect unauthorized data persistence and access findings to detect the targeted failure.

B:

The proposed measure would suit a setting where an agent reads content written by unknown third parties, since the control layer can enumerate trust-boundary crossings and test tool redirection attempts; by contrast, the energy operations release needs evidence from unauthorized data persistence and access findings.

C:

When new prompt versions or changing input mix alter policy compliance, the control layer can establish post-release monitoring and re-evaluation triggers; however, that proxy would not demonstrate unauthorized data persistence and access findings for the selected intervention.

D:

Where an assistant states confident conclusions without revealing evidence limitations, a relevant check would establish whether the control layer can require source grounding and explicit uncertainty when proof is unavailable; but the acceptance gate must instead inspect unauthorized data persistence and access findings to detect the targeted failure.

Question 341

During an acceptance review in media licensing, Parallel independent retrieval is operational but its benefit has not been measured. Auditors are concerned that sequential calls inflate end-to-end wait time without improving correctness. Which evidence most directly settles the question?

  1. Measure truthful-defer rate and user task completion after handoff
  2. Measure error reduction versus added p95 latency by risk tier
  3. Measure quality versus latency across task difficulty buckets
  4. Measure critical-path retrieval latency and source coverage

Explanation

Correct Answer

D:

The release should evaluate critical-path retrieval latency and source coverage rather than equate successful calls with real outcomes. The control works on the premise that the integration can fetch independent sources concurrently and join their authorized results; the proposed measure can expose the risk that sequential calls inflate end-to-end wait time without improving correctness.

Incorrect Answers

A:

Where a strict response-time target sometimes conflicts with deep evidence searches, a relevant check would establish whether the integration can provide an explicit partial or queued result rather than fabricate a quick full answer; by contrast, the media licensing release needs evidence from critical-path retrieval latency and source coverage.

B:

When routine descriptions are low stakes but payment changes require evidence checking, the integration can apply expensive verification only to high-impact or low-confidence actions; however, that proxy would not demonstrate critical-path retrieval latency and source coverage for the selected intervention.

C:

When straightforward lookup requests, deep analysis use very different effort and the integration can allocate larger reasoning budgets only to tasks that need them; but the acceptance gate must instead inspect critical-path retrieval latency and source coverage to detect the targeted failure.

Question 342

A telecommunications support release candidate fails review after assessors observe that manual spot checks miss parser-breaking regressions. What intervention should take priority?

  1. Introduce Automated contract test suite
  2. Introduce Human blinded rating panel
  3. Introduce Adversarial injection fixtures
  4. Introduce Metamorphic consistency tests

Explanation

Correct Answer

A:

This intervention addresses the failure at its source because the evaluation process can validate outputs mechanically on every release candidate. To keep the issue from recurring, establish frozen consumer contract and test harness and verify schema pass rate and malformed payload count; otherwise manual spot checks miss parser-breaking regressions can persist.

Incorrect Answers

B:

Although the alternative reduces the risk that knowing model identity can bias judgments of equivalent outputs because the evaluation process can collect independent reviewer scores without revealing model or variant; yet it would not correct the observed weakness: manual spot checks miss parser-breaking regressions.

C:

When normal benign examples do not exercise trust boundaries, the evaluation process can add attacker-controlled content and escalation attempts to evaluation; but this remedy operates on another boundary, not the case where manual spot checks miss parser-breaking regressions.

D:

This addresses the failure where a single fixed example cannot expose brittle prompt sensitivity, because the evaluation process can apply controlled input transformations and compare invariant outcomes; whereas the telecommunications support defect follows the path in which manual spot checks miss parser-breaking regressions.

Question 343

During solution discovery in media licensing, a generated instruction could trigger a forbidden account mutation. Which proposed approach handles this situation at its source?

  1. Audited output-normalization stage
  2. Role and instruction separation
  3. Explicit schema and refusal conditions
  4. Guardrail outside generation

Explanation

Correct Answer

D:

The architectural distinction is that the model configuration can enforce authorization and policy in deterministic application controls. The decision should be implemented with server-side policy enforcement before side effects and evaluated against prevented unauthorized actions during adversarial tests; otherwise, a warning in the prompt does not prevent a tool executor from making a prohibited change.

Incorrect Answers

A:

When downstream clients need stable currency and date conventions, the model configuration can normalize and validate final fields after model generation; it does not ensure the required behavior: the model configuration can enforce authorization and policy in deterministic application controls.

B:

When retrieved customer text contains hostile instructions to override policy, the model configuration can place durable task rules in trusted instruction channels and treat retrieved text as data; whereas the media licensing case must guard against the outcome that a warning in the prompt does not prevent a tool executor from making a prohibited change.

C:

For a workload where the caller needs a machine-readable response even when evidence is insufficient, the model configuration can define output fields and abstention behavior for unanswerable inputs; the alternative leaves the central requirement unproven: the model configuration can enforce authorization and policy in deterministic application controls.

Question 344

The human-resources services staging review finds that undifferentiated feedback cannot guide targeted changes. Which design intervention is the most appropriate corrective action?

  1. Prioritize Idempotent action boundary
  2. Prioritize Typed output validation and repair route
  3. Prioritize Explicit intermediate state store
  4. Prioritize Feedback taxonomy with ground-truth capture

Explanation

Correct Answer

D:

The reported failure calls for the control under which structured error tags make future fixes attributable to retrieval, prompt or policy causes. A prompt-only reminder or unrelated metric would not address the risk that undifferentiated feedback cannot guide targeted changes. Evidence of resolution requires trained reviewers and a stable reason-code dictionary and an improvement in share of reviewed errors with actionable classification.

Incorrect Answers

A:

Where a retrying agent sometimes creates the same support task twice, idempotency keys and deduplication protect side-effecting writes across retries; but this remedy operates on another boundary, not the case where undifferentiated feedback cannot guide targeted changes.

B:

This addresses the failure where accepting plausible-looking prose can break consumers or misroute cases, because the architecture can validate generated payloads against a contract and route failures to repair or review; whereas the human-resources services defect follows the path in which undifferentiated feedback cannot guide targeted changes.

C:

This addresses the failure where relying on conversation text alone can duplicate side effects after failure, because durable state records completed stages and enables safe continuation; yet it would not correct the observed weakness: undifferentiated feedback cannot guide targeted changes.

Question 345

An independent assessment of a supply-chain planning Claude system reports that sequential calls inflate end-to-end wait time without improving correctness. Which change best closes the exposed control gap?

  1. Deploy Asynchronous completion for long tasks
  2. Deploy Parallel independent retrieval
  3. Deploy Rerank only the candidate set
  4. Deploy Selective verification for risky outputs

Explanation

Correct Answer

B:

This is the causal correction: the integration can fetch independent sources concurrently and join their authorized results. With independent retrievers and bounded parallelism in place, the team can test critical-path retrieval latency and source coverage and show that the problematic case in which sequential calls inflate end-to-end wait time without improving correctness has been addressed.

Incorrect Answers

A:

When holding an interactive connection open invites timeouts and duplicate retries, the integration can submit a job and return an identifier with status tracking; whereas the supply-chain planning defect follows the path in which sequential calls inflate end-to-end wait time without improving correctness.

C:

Where broad retrieval finds relevant content but top results contain noisy matches, the integration can rerank a bounded candidate list for precision before generation; but this remedy operates on another boundary, not the case where sequential calls inflate end-to-end wait time without improving correctness.

D:

This addresses the failure where uniform verification can miss the interactive latency objective, because the integration can apply expensive verification only to high-impact or low-confidence actions; whereas the supply-chain planning defect follows the path in which sequential calls inflate end-to-end wait time without improving correctness.

Question 346

Following a regression in media licensing, the team establishes that generic assurances about safety do not settle operational accountability. Which proposed redesign most directly prevents recurrence?

  1. Apply Risk-and-mitigation narrative
  2. Apply Architecture decision record
  3. Apply TCO and change-cost briefing
  4. Apply Explicit assumption register

Explanation

Correct Answer

A:

The problem lies in the absence of a mechanism where the delivery team can explain residual risk, enforceable mitigations and explicit escalation boundaries. Reinstating it, backed by common risk taxonomy and mitigation evidence, limits the exposure that generic assurances about safety do not settle operational accountability; the engineering team should track signed risk acceptance and validated control ownership.

Incorrect Answers

B:

When informal chat history loses rationale and assumed constraints, the delivery team can record context, alternatives, decision and consequences in a versioned ADR; but this remedy operates on another boundary, not the case where generic assurances about safety do not settle operational accountability.

C:

When inference-only estimates omit significant lifecycle costs, the delivery team can include maintenance, migration, vendor dependencies and personnel effort in comparison; whereas the media licensing defect follows the path in which generic assurances about safety do not settle operational accountability.

D:

When hidden premises become expensive surprises during rollout, the delivery team can track assumptions, validation owner and consequences of failure; yet it would not correct the observed weakness: generic assurances about safety do not settle operational accountability.

Question 347

A financial operations pilot cannot be accepted because encryption in transit alone does not resolve overcollection or retention. The owner can change one primary architectural control; what should it be?

  1. Enforce Data exposure and retention review
  2. Enforce Tool-chain compromise assessment
  3. Enforce Prompt-injection threat modeling
  4. Enforce Model behavior drift surveillance
  5. Enforce Overclaim and opacity mitigation

Explanation

Correct Answer

A:

The remediation acts at the relevant design boundary: the control layer can trace data paths, access, retention and deletion obligations before launch. That counteracts the observed hazard that encryption in transit alone does not resolve overcollection or retention. Rollout should include approved data-flow inventory and service agreements, with regression evidence from unauthorized data persistence and access findings.

Incorrect Answers

B:

This addresses the failure where security of the model endpoint alone does not secure every tool hop, because the control layer can identify propagation paths for compromised tokens and manipulated outputs; whereas the financial operations defect follows the path in which encryption in transit alone does not resolve overcollection or retention.

C:

Although the alternative reduces the risk that treating retrieved instructions as harmless text overlooks agentic abuse because the control layer can enumerate trust-boundary crossings and test tool redirection attempts; yet it would not correct the observed weakness: encryption in transit alone does not resolve overcollection or retention.

D:

Although the alternative reduces the risk that a one-time approval cannot guarantee future behavior across drift because the control layer can establish post-release monitoring and re-evaluation triggers; but this remedy operates on another boundary, not the case where encryption in transit alone does not resolve overcollection or retention.

E:

Where an assistant states confident conclusions without revealing evidence limitations, the control layer can require source grounding and explicit uncertainty when proof is unavailable; whereas the financial operations defect follows the path in which encryption in transit alone does not resolve overcollection or retention.

Question 348

A travel operations audit challenges the need for Automated contract test suite. Which risk would remain unmanaged if its function were not provided elsewhere?

  1. A failure where a single fixed example cannot expose brittle prompt sensitivity
  2. A failure where manual spot checks miss parser-breaking regressions
  3. A failure where a high average score can mask catastrophic minority failures
  4. A failure where static snapshots fail to reveal stale-index behavior
  5. A failure where normal benign examples do not exercise trust boundaries

Explanation

Correct Answer

B:

This is the direct residual hazard: manual spot checks miss parser-breaking regressions. The omitted intervention matters because the evaluation process can validate outputs mechanically on every release candidate. A reviewer should demand frozen consumer contract and test harness and measure schema pass rate and malformed payload count before treating the risk as controlled.

Incorrect Answers

A:

If answers should not change when irrelevant formatting or names change, a single fixed example cannot expose brittle prompt sensitivity; however, the named safeguard addresses the distinct hazard that manual spot checks miss parser-breaking regressions.

C:

The other exposure arises when long-tail failure classes rarely appear in a random convenience sample: a high average score can mask catastrophic minority failures. Omitting the requested control instead leaves the more immediate risk that manual spot checks miss parser-breaking regressions; but this is not the immediate loss associated with omitting Automated contract test suite: manual spot checks miss parser-breaking regressions.

D:

When policy retrieval quality looks good using documents from the test period, static snapshots fail to reveal stale-index behavior; however, the named safeguard addresses the distinct hazard that manual spot checks miss parser-breaking regressions.

E:

If a tool agent ingests external documents that can contain hostile instructions, normal benign examples do not exercise trust boundaries; whereas removing Automated contract test suite primarily exposes the risk that manual spot checks miss parser-breaking regressions.

Question 349

An education-services team passed its happy-path demo, and a reviewer asks to discard Hook-based policy checks. Which problem could reappear even after a successful demo?

  1. An exposure in which unrestricted execution can alter secrets or infrastructure accidentally
  2. An exposure in which a shared default environment increases cross-project exposure
  3. An exposure in which natural-language reminders cannot substitute for enforceable test execution
  4. An exposure in which ad hoc personal connectors produce inconsistent and insecure access

Explanation

Correct Answer

C:

The loss of the control permits the situation where natural-language reminders cannot substitute for enforceable test execution. Its governing mechanism is that the engineering workflow can run deterministic checks at controlled workflow boundaries; without an equivalent mechanism, a favorable demonstration cannot establish safety or quality.

Incorrect Answers

A:

If coding agents request shell access beyond the repository task, unrestricted execution can alter secrets or infrastructure accidentally; whereas removing Hook-based policy checks primarily exposes the risk that natural-language reminders cannot substitute for enforceable test execution.

B:

The competing hazard is that a shared default environment increases cross-project exposure in a workload where confidential projects share agent context and local tool access inadvertently; but this is not the immediate loss associated with omitting Hook-based policy checks: natural-language reminders cannot substitute for enforceable test execution.

D:

When development teams connect tools with overlapping credentials and contracts, ad hoc personal connectors produce inconsistent and insecure access; whereas removing Hook-based policy checks primarily exposes the risk that natural-language reminders cannot substitute for enforceable test execution.

Question 350

A financial research assistant now performs two independent evidence lookups sequentially even though neither depends on the other. What performance consequence comes from dropping concurrent retrieval?

  1. A failure where a fixed high budget can waste time and tokens on simple cases
  2. A failure where holding an interactive connection open invites timeouts and duplicate retries
  3. A failure where cutting retrieval silently may produce unsupported definitive statements
  4. A failure where sequential calls inflate end-to-end wait time without improving correctness

Explanation

Correct Answer

D:

The relevant failure mode is that sequential calls inflate end-to-end wait time without improving correctness. The existing design can prevent it when the integration can fetch independent sources concurrently and join their authorized results. The team needs independent retrievers and bounded parallelism and must monitor critical-path retrieval latency and source coverage to support any claim of equivalent protection.

Incorrect Answers

A:

If straightforward lookup requests and deep analysis use very different effort, a fixed high budget can waste time and tokens on simple cases; but this is not the immediate loss associated with omitting Parallel independent retrieval: sequential calls inflate end-to-end wait time without improving correctness.

B:

When complex report synthesis exceeds an interactive HTTP timeout, holding an interactive connection open invites timeouts and duplicate retries; however, the named safeguard addresses the distinct hazard that sequential calls inflate end-to-end wait time without improving correctness.

C:

When a strict response-time target sometimes conflicts with deep evidence searches, cutting retrieval silently may produce unsupported definitive statements; whereas removing Parallel independent retrieval primarily exposes the risk that sequential calls inflate end-to-end wait time without improving correctness.

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