Anthropic CCAR-P (Claude Certified Architect - Professional) Practice Test 3

Question 20

An operations analyst for enterprise knowledge management must evaluate Small-model first-pass routing on real tasks. A test must detect whether always using the strongest model increases cost and wait time, even when outputs appear fluent. Which measure is appropriate?

  1. Audit cost-quality-latency Pareto frontier on held-out examples
  2. Review audited synthesis accuracy on difficult cases
  3. Audit safe fallback completion rate under simulated outages
  4. Audit routed quality at the target cost and p95 latency

Explanation

Correct Answer

D:

Tracking routed quality at the target cost and p95 latency provides direct evidence that the system delivers its intended effect: a fast adequate tier handles easy cases and escalates only uncertain ones. The measure is more informative than prompt counts because the failure being measured is that always using the strongest model increases cost and wait time.

Incorrect Answers

A:

Where stakeholders choose a model based on brand rather than target workload, a relevant check would establish whether the model configuration can evaluate candidate models on identical task data and operating constraints; by contrast, the enterprise knowledge management release needs evidence from routed quality at the target cost and p95 latency.

B:

The proposed measure would suit a setting where multi-document arbitration requires several interdependent inferences, since a stronger model can justify tradeoffs across heterogeneous evidence; however, that proxy would not demonstrate routed quality at the target cost and p95 latency for the selected intervention.

C:

When primary inference service can time out during a critical support window, the model configuration can route to an approved fallback model with compatible output contracts; but the acceptance gate must instead inspect routed quality at the target cost and p95 latency to detect the targeted failure.

Question 21

A travel operations incident exposes a specific weakness: feature demonstrations can obscure the absence of an accountable success condition. Which architectural correction addresses the cause rather than masking the symptom?

  1. Prioritize Feasibility pilot with representative cases
  2. Prioritize Outcome-first use-case framing
  3. Prioritize Human-approved recommendation service
  4. Prioritize Value-to-risk prioritization

Explanation

Correct Answer

B:

This is the causal correction: a measurable business outcome narrows the solution boundary before model selection. With an agreed baseline and accountable business owner in place, the team can test change in qualified escalations per thousand completed tasks and show that the problematic case in which feature demonstrations can obscure the absence of an accountable success condition has been addressed.

Incorrect Answers

A:

Where the proposed automation encounters exceptions that have never been tested, small-scope trials expose feasibility and exception handling before a broad commitment; whereas the travel operations defect follows the path in which feature demonstrations can obscure the absence of an accountable success condition.

C:

Although the alternative reduces the risk that automating the last step could bypass required accountability because Claude drafts ranked recommendations while a designated reviewer authorizes each action; but this remedy operates on another boundary, not the case where feature demonstrations can obscure the absence of an accountable success condition.

D:

Where candidate use cases differ in savings, uncertainty and personal-data exposure, a scored decision framework surfaces high-value lower-risk starting points; whereas the travel operations defect follows the path in which feature demonstrations can obscure the absence of an accountable success condition.

Question 22

A research administration implementation team has approved Repository-level project instructions. Its expected behavior is that the engineering workflow can version shared Claude Code guidance alongside the project and review changes. What prerequisite belongs on the release-readiness checklist?

  1. Arrange team permission policy and escalation route
  2. Arrange authorized hooks and maintained check commands
  3. Arrange isolated identities and explicit workspace inventory
  4. Arrange approved project guidance and code review owners

Explanation

Correct Answer

D:

Readiness depends on approved project guidance and code review owners, not just a design approval. That condition supports the mechanism in which the engineering workflow can version shared Claude Code guidance alongside the project and review changes. Without it, personal chat instructions do not establish reproducible team conventions; monitoring standards adherence across fresh developer workspaces helps verify the deployed result.

Incorrect Answers

A:

The setup is useful to enable the alternative's behavior where the engineering workflow can configure approved actions and deny risky operations by default and coding agents request shell access beyond the repository task; however, the named intervention needs approved project guidance and code review owners before its protection can be enforced.

B:

For the separate challenge where code generation sometimes bypasses mandatory lint and secret scans, the engineering workflow can run deterministic checks at controlled workflow boundaries; but the production dependency for Repository-level project instructions is approved project guidance and code review owners.

C:

For the separate challenge where confidential projects share agent context and local tool access inadvertently, the engineering workflow can isolate project roots, secrets and tool permissions by approved workspace; for this release, the team must first verify approved project guidance and code review owners.

Question 23

An independent assessment of a procurement services Claude system reports that generic satisfaction scores may reward confident errors. Which change best closes the exposed control gap?

  1. Prioritize Calibration and abstention quality
  2. Prioritize Task-level acceptance rubric
  3. Prioritize Tail-latency measurement
  4. Prioritize Grounded-claim verification rate

Explanation

Correct Answer

B:

This intervention addresses the failure at its source because the evaluation process can define observable criteria for a usable output rather than aesthetic fluency. To keep the issue from recurring, establish agreed pass-fail examples and acceptance threshold and verify expert-reviewed acceptance rate by task category; otherwise generic satisfaction scores may reward confident errors can persist.

Incorrect Answers

A:

Where incomplete evidence should yield escalation rather than invented certainty, the evaluation process can evaluate correctness jointly with justified abstentions across uncertainty levels; but this remedy operates on another boundary, not the case where generic satisfaction scores may reward confident errors.

C:

Where interactive sessions frequently exceed their response deadline, the evaluation process can track p95 and p99 end-to-end latency by request class; yet it would not correct the observed weakness: generic satisfaction scores may reward confident errors.

D:

Where the system summarizes regulated policies with citations, the evaluation process can score whether factual claims are supported by cited source spans; but this remedy operates on another boundary, not the case where generic satisfaction scores may reward confident errors.

Question 24

A financial operations team is selecting the next improvement for its AI service; a complex connector ecosystem cannot fit full tool descriptions in each request. What design choice has the strongest direct justification?

  1. Adopt Task-scoped tool allowlist
  2. Adopt Progressive tool discovery
  3. Adopt Tool ownership and lifecycle review
  4. Adopt Capped tool-loop execution

Explanation

Correct Answer

B:

The architectural distinction is that the integration can advertise compact discovery metadata and load detailed definitions on demand. The decision should be implemented with curated discoverability metadata and scoped authorization and evaluated against tool-definition tokens per resolved user task; otherwise, eager inclusion of the entire catalog wastes context and confuses selection.

Incorrect Answers

A:

When a research agent is offered dozens of irrelevant mutation tools, the integration can expose only tools needed for the active task and user permission scope; the stated scenario depends on a different mechanism: the integration can advertise compact discovery metadata and load detailed definitions on demand.

C:

For a workload where several plugins expose overlapping outdated customer lookup functions, the integration can retire or consolidate redundant capabilities with a responsible owner; whereas the financial operations case must guard against the outcome that eager inclusion of the entire catalog wastes context and confuses selection.

D:

When a search agent endlessly retries ineffective tools on malformed identifiers, the integration can enforce retry, token and call budgets with a deliberate stop path; the stated scenario depends on a different mechanism: the integration can advertise compact discovery metadata and load detailed definitions on demand.

Question 25

An insurance processing pilot cannot be accepted because executive summaries alone omit edge cases that determine project success. The owner can change one primary architectural control; what should it be?

  1. Deploy Data readiness assessment
  2. Deploy Workflow observation interviews
  3. Deploy Stakeholder and decision-rights mapping
  4. Deploy Functional versus nonfunctional split

Explanation

Correct Answer

B:

The reported failure calls for the control under which the delivery team can observe actual exception handling and frontline work before mapping automation. A prompt-only reminder or unrelated metric would not address the risk that executive summaries alone omit edge cases that determine project success. Evidence of resolution requires access to operators and representative case samples and an improvement in discovered exceptions covered by the agreed requirements.

Incorrect Answers

A:

When a strong model cannot compensate for absent or unauthorized source data, the delivery team can validate quality, permissions, update cadence and ownership before promising RAG results; yet it would not correct the observed weakness: executive summaries alone omit edge cases that determine project success.

C:

The alternative reduces the risk that unclear authority causes late redesign and unowned risk decisions because the delivery team can define accountable decision owners and consultation rights early; however, the insurance processing defect follows the path in which executive summaries alone omit edge cases that determine project success.

D:

Where a proposal lists features but omits response time and recovery requirements, the delivery team can capture distinct behavior, latency, security, availability and data constraints; yet it would not correct the observed weakness: executive summaries alone omit edge cases that determine project success.

Question 26

The post-incident analysis for energy operations concludes that prompt-only guidance cannot neutralize malicious external text. Which remediation targets the relevant failure boundary?

  1. Deploy Output privacy filter
  2. Deploy Ingress content classification
  3. Deploy Untrusted retrieved-content isolation
  4. Deploy Tool-execution policy gate

Explanation

Correct Answer

B:

This is the causal correction: the control layer can classify untrusted inputs and prevent them from acquiring instruction authority. With documented trust boundaries and security test cases in place, the team can test prompt-injection success rate against controlled attack fixtures and show that the problematic case in which prompt-only guidance cannot neutralize malicious external text has been addressed.

Incorrect Answers

A:

When upstream input minimization alone may not catch derived disclosures, the control layer can enforce recipient-specific redaction and validated output fields; but this remedy operates on another boundary, not the case where prompt-only guidance cannot neutralize malicious external text.

C:

Where supplier documents include embedded instructions masquerading as system messages, the control layer can treat retrieved text as evidence and prohibit it from changing tool policy; yet it would not correct the observed weakness: prompt-only guidance cannot neutralize malicious external text.

D:

This addresses the failure where model self-restraint cannot guarantee that a bad action is blocked, because the control layer can require server-side authorization and business-policy checks before each side effect; but this remedy operates on another boundary, not the case where prompt-only guidance cannot neutralize malicious external text.

Question 27

An education-services team passed its happy-path demo, and a reviewer asks to discard Outcome-first use-case framing. Which concern should the approval board raise?

  1. A failure where a portfolio of speculative features can crowd out working high-value uses
  2. A failure where a polished demo can hide unacceptable long-tail failures
  3. A failure where feature demonstrations can obscure the absence of an accountable success condition
  4. A failure where deployment can appear successful while adding hidden review effort

Explanation

Correct Answer

C:

If the protection disappears, feature demonstrations can obscure the absence of an accountable success condition. The missing architectural distinction is that a measurable business outcome narrows the solution boundary before model selection. The approval board should require an alternative with an agreed baseline and accountable business owner and validate it against change in qualified escalations per thousand completed tasks.

Incorrect Answers

A:

The competing hazard is that a portfolio of speculative features can crowd out working high-value uses in a workload where a team wants to fund a large platform before validating which use case helps; however, the named safeguard addresses the distinct hazard that feature demonstrations can obscure the absence of an accountable success condition.

B:

The other exposure arises when the proposed automation encounters exceptions that have never been tested: a polished demo can hide unacceptable long-tail failures; whereas removing Outcome-first use-case framing primarily exposes the risk that feature demonstrations can obscure the absence of an accountable success condition.

D:

The other exposure arises when executives claim the assistant will save time without measuring current work: deployment can appear successful while adding hidden review effort; however, the named safeguard addresses the distinct hazard that feature demonstrations can obscure the absence of an accountable success condition.

Question 28

A media licensing incident exposes a specific weakness: always using the strongest model increases cost and wait time. Which architectural correction addresses the cause rather than masking the symptom?

  1. Introduce Small-model first-pass routing
  2. Introduce Reasoning-capable model for complex synthesis
  3. Introduce Context-capacity fit before model selection
  4. Introduce Per-class model policy

Explanation

Correct Answer

A:

The remediation acts at the relevant design boundary: a fast adequate tier handles easy cases and escalates only uncertain ones. That counteracts the observed hazard that always using the strongest model increases cost and wait time. Rollout should include measured task complexity and escalation thresholds, with regression evidence from routed quality at the target cost and p95 latency.

Incorrect Answers

B:

Although the alternative reduces the risk that a lightweight classifier can omit crucial cross-document reasoning because a stronger model can justify tradeoffs across heterogeneous evidence; but this remedy operates on another boundary, not the case where always using the strongest model increases cost and wait time.

C:

The alternative reduces the risk that nominal context limits alone do not prove reliable use of every passage because the model configuration can test effective context utilization and recall before choosing a model configuration; however, the media licensing defect follows the path in which always using the strongest model increases cost and wait time.

D:

Where privacy-sensitive and low-risk workloads require different operating controls, the model configuration can set explicit allowed model configurations by data class and purpose; yet it would not correct the observed weakness: always using the strongest model increases cost and wait time.

Question 29

The contract administration architect proposes omitting Task-level acceptance rubric from the production design. What is the most direct failure mode that decision would expose?

  1. An exposure in which citation formatting alone does not prove support for the claim
  2. An exposure in which accuracy on answered cases alone can reward refusing everything
  3. An exposure in which generic satisfaction scores may reward confident errors
  4. An exposure in which median latency can hide the failures users notice most
  5. An exposure in which raw tokens per request omit cost of corrections and failures

Explanation

Correct Answer

C:

This is the direct residual hazard: generic satisfaction scores may reward confident errors. The omitted intervention matters because the evaluation process can define observable criteria for a usable output rather than aesthetic fluency. A reviewer should demand agreed pass-fail examples and acceptance threshold and measure expert-reviewed acceptance rate by task category before treating the risk as controlled.

Incorrect Answers

A:

The competing hazard is that citation formatting alone does not prove support for the claim in a workload where the system summarizes regulated policies with citations; but this is not the immediate loss associated with omitting Task-level acceptance rubric: generic satisfaction scores may reward confident errors.

B:

When incomplete evidence should yield escalation rather than invented certainty, accuracy on answered cases alone can reward refusing everything; however, the named safeguard addresses the distinct hazard that generic satisfaction scores may reward confident errors.

D:

The other exposure arises when interactive sessions frequently exceed their response deadline: median latency can hide the failures users notice most. Omitting the requested control instead leaves the more immediate risk that generic satisfaction scores may reward confident errors; but this is not the immediate loss associated with omitting Task-level acceptance rubric: generic satisfaction scores may reward confident errors.

E:

The competing hazard is that raw tokens per request omit cost of corrections and failures in a workload where model usage drops while human rework unexpectedly grows; however, the named safeguard addresses the distinct hazard that generic satisfaction scores may reward confident errors.

Question 30

The field-service coordination go-live board accepts the premise that the integration can advertise compact discovery metadata and load detailed definitions on demand. Which preparation is indispensable before it can rely on Progressive tool discovery operationally?

  1. Confirm inventory with owner, version and retirement process
  2. Confirm curated discoverability metadata and scoped authorization
  3. Confirm loop policy and classified tool-error responses
  4. Confirm documented action schemas and server-side validation
  5. Confirm distinct service identities and audited write approvals
  6. Confirm owner-approved capability mapping per task

Explanation

Correct Answer

B:

The prerequisite creates the necessary operating boundary: curated discoverability metadata and scoped authorization. The system is intended to work because the integration can advertise compact discovery metadata and load detailed definitions on demand. A release without that foundation may still suffer eager inclusion of the entire catalog wastes context and confuses selection; use tool-definition tokens per resolved user task to evaluate the result.

Incorrect Answers

A:

For the separate challenge where several plugins expose overlapping outdated customer lookup functions, the integration can retire or consolidate redundant capabilities with a responsible owner; for this release, the team must first verify curated discoverability metadata and scoped authorization.

C:

This prerequisite supports a method used when a search agent endlessly retries ineffective tools on malformed identifiers, because the integration can enforce retry, token and call budgets with a deliberate stop path; but the production dependency for Progressive tool discovery is curated discoverability metadata and scoped authorization.

D:

Where a generic execute-any-query tool lets a model submit unreviewed commands, the integration can replace broad verbs with narrow parameterized operations and validation; for this release, the team must first verify curated discoverability metadata and scoped authorization.

E:

For the separate challenge where an assistant must read audit events but rarely modify audit settings, the integration can expose read-only access independently from privileged changes; however, the named intervention needs curated discoverability metadata and scoped authorization before its protection can be enforced.

F:

Where a research agent is offered dozens of irrelevant mutation tools, the integration can expose only tools needed for the active task and user permission scope; but the production dependency for Progressive tool discovery is curated discoverability metadata and scoped authorization.

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