Anthropic CCAR-P (Claude Certified Architect - Professional) Practice Test 91

Question 900

An education services pilot cannot be accepted because selecting the nearest vector does not make it relevant. The owner can change one primary architectural control; what should it be?

  1. Introduce Time-aware policy retrieval
  2. Introduce Exact database query through a typed tool
  3. Introduce Selective SQL plus document retrieval
  4. Introduce Relevance-threshold abstention

Explanation

Correct Answer

D:

The reported failure calls for the control under which the integration can require evidence strength before generating a confident factual answer. A prompt-only reminder or unrelated metric would not address the risk that selecting the nearest vector does not make it relevant. Evidence of resolution requires approved enabling configuration and accountable owner and an improvement in unsupported-answer refusal rate and missed-answer rate.

Incorrect Answers

A:

The alternative reduces the risk that always using the newest document can answer a historical question incorrectly because the integration can filter versions by effective interval before selecting passages; however, the education services defect follows the path in which selecting the nearest vector does not make it relevant.

B:

Although the alternative reduces the risk that generative similarity search is unsuitable for exact transactional state because the integration can retrieve authorized structured records using parameterized queries; yet it would not correct the observed weakness: selecting the nearest vector does not make it relevant.

C:

When either retrieval source alone omits one class of necessary evidence, the integration can join exact database results with grounded policy excerpts in the application; but this remedy operates on another boundary, not the case where selecting the nearest vector does not make it relevant.

Question 901

The architecture review for media licensing has this operational target: the architecture can summarize bounded chunks then aggregate their structured findings. Which THREE choices provide the solution, its enabling condition, and a defensible release measure? Choose THREE.

  1. Implement Map-reduce summarization
  2. Implement Exception-first task splitting
  3. Track conflicts detected and resolved with provenance
  4. Track retained key-fact recall and final contradiction rate
  5. Establish stable chunk boundaries and merge rules
  6. Establish declared dependencies and persistent node state

Explanation

Correct Answers

A:

The architecture component is justified by its behavior: the architecture can summarize bounded chunks then aggregate their structured findings. It targets the exposure that concatenating every note increases cost and can truncate important evidence, while the complementary selected actions provide readiness and, for a three-part package, evidence.

D:

This is an acceptance signal rather than another design control: retained key-fact recall and final contradiction rate. Evaluating that signal on representative tasks tests whether the architecture can summarize bounded chunks then aggregate their structured findings and identifies cases where concatenating every note increases cost and can truncate important evidence.

E:

Before rollout, the team needs stable chunk boundaries and merge rules. This is a deployment dependency for the intervention whose effect is that the architecture can summarize bounded chunks then aggregate their structured findings, not an alternative metric or unrelated design feature.

Incorrect Answers

B:

This design choice is aimed at cases where typical requests succeed but rare regulatory exceptions cause serious errors; it is effective because the architecture can separate normal and exception branches with stronger checks for the latter; however, the intended intervention requires that the architecture can summarize bounded chunks then aggregate their structured findings.

C:

This test would validate a different benefit: the architecture can separate conflict detection and resolution from answer drafting; but that indicator does not verify retained key-fact recall and final contradiction rate for the selected architecture.

F:

The alternative setup enables a different mechanism: the architecture can model the task as a DAG and execute nodes only when prerequisites are satisfied; whereas this intervention requires stable chunk boundaries and merge rules as its operating precondition.

Question 902

A healthcare operations service passed its happy-path demo, and a reviewer asks to discard Relevance-threshold abstention. Which concern should the approval board raise?

  1. The possibility that always using the newest document can answer a historical question incorrectly
  2. The possibility that vector neighbors alone do not prove multi-hop entity relationships
  3. The possibility that a single literal query can miss synonym-heavy material
  4. The possibility that generative similarity search is unsuitable for exact transactional state
  5. The possibility that selecting the nearest vector does not make it relevant

Explanation

Correct Answer

E:

Removing the safeguard exposes a failure in which selecting the nearest vector does not make it relevant. The architecture normally counters this because the integration can require evidence strength before generating a confident factual answer. A compensating design would require approved enabling configuration and accountable owner and evidence from unsupported-answer refusal rate and missed-answer rate.

Incorrect Answers

A:

When users ask what rule applied on a particular historical date, always using the newest document can answer a historical question incorrectly; but this is not the immediate loss associated with omitting Relevance-threshold abstention: selecting the nearest vector does not make it relevant.

B:

The other exposure arises when investigators ask how two entities connect through approved relationships: vector neighbors alone do not prove multi-hop entity relationships; however, the named safeguard addresses the distinct hazard that selecting the nearest vector does not make it relevant.

C:

The competing hazard is that a single literal query can miss synonym-heavy material in a workload where relevant policy is difficult to find because employees use many aliases; whereas removing Relevance-threshold abstention primarily exposes the risk that selecting the nearest vector does not make it relevant.

D:

When customers ask for a current balance identified by account ID, generative similarity search is unsuitable for exact transactional state; but this is not the immediate loss associated with omitting Relevance-threshold abstention: selecting the nearest vector does not make it relevant.

Question 903

A human-resources services team is selecting the next improvement for its AI service; three independent classifications currently run sequentially. What design choice has the strongest direct justification?

  1. Introduce Parallelizable workflow optimization
  2. Introduce Result-driven model escalation
  3. Introduce Optimization through better task success
  4. Introduce Adaptive tool timeout budget

Explanation

Correct Answer

A:

The intervention acts on the system behavior directly: the evaluation process can parallelize work while joining outputs under a bounded concurrency cap. Before rollout, verify independence proof and capacity testing. The relevant success test is overall task latency and quota error rate; otherwise blindly increasing concurrency can overload downstream quotas.

Incorrect Answers

B:

When lightweight answers work on simple cases but fail specialized regulatory reasoning, the evaluation process can raise model capacity selectively after evidence-based difficulty checks; it does not ensure the required behavior: the evaluation process can parallelize work while joining outputs under a bounded concurrency cap.

C:

When shaving tokens degrades downstream extraction precision, the evaluation process can compare token savings against full correction and review cost; whereas the human-resources services case must guard against the outcome that blindly increasing concurrency can overload downstream quotas.

D:

When a task waits for a slow optional analytics integration, the evaluation process can use per-dependency deadlines and continue only if required evidence is available; the alternative leaves the central requirement unproven: the evaluation process can parallelize work while joining outputs under a bounded concurrency cap.

Question 904

The travel operations deployment uses Prompt and model configuration manifest, but executive reports currently count only successful model responses. The residual concern is that source-control of application code alone does not recreate inference behavior. Which acceptance measure would reveal real progress?

  1. Measure requirements with assigned controls and passing evidence
  2. Measure environment parity and reproducible rollback
  3. Measure contract test pass rate across consumer versions
  4. Measure recovery drill success and time to restore service

Explanation

Correct Answer

B:

The acceptance measure focuses on environment parity and reproducible rollback at the task or workflow boundary. Model fluency would not establish whether the delivery team can version prompts, model choices, tools and safety settings together, and it could conceal cases where source-control of application code alone does not recreate inference behavior.

Incorrect Answers

A:

The proposed measure would suit a setting where stakeholders cannot tell which architectural control satisfies which need, since the delivery team can link decisions to acceptance tests and authoritative requirement IDs; but the acceptance gate must instead inspect environment parity and reproducible rollback to detect the targeted failure.

C:

The proposed measure would suit a setting where separate teams implement tool providers and consumers inconsistently, since the delivery team can publish request schemas, response semantics, error categories and retries; however, that proxy would not demonstrate environment parity and reproducible rollback for the selected intervention.

D:

The proposed measure would suit a setting where incident responders rely on the original author for every production failure, since the delivery team can document symptoms, safe mitigation, escalation and rollback procedures; but the acceptance gate must instead inspect environment parity and reproducible rollback to detect the targeted failure.

Question 905

After installing Data residency deployment review in a contract administration workflow, the team must validate whether it has reduced the exposure that selecting one regional database does not constrain every connected service. Which measurement is most relevant?

  1. Track vendor agreement coverage and PHI exposure incidents
  2. Track audit evidence completeness across sampled cases
  3. Track forbidden region transfers found in data-flow audits
  4. Track documented lawful-basis coverage and purpose exceptions

Explanation

Correct Answer

C:

Tracking forbidden region transfers found in data-flow audits provides direct evidence that the system delivers its intended effect: the control layer can verify regional processing, logs, backups and support access as a whole. The measure is more informative than prompt counts because the failure being measured is that selecting one regional database does not constrain every connected service.

Incorrect Answers

A:

Where a US clinical workflow handles protected health information through vendors, a relevant check would establish whether the control layer can verify applicable HIPAA obligations and required business associate arrangements; by contrast, the contract administration release needs evidence from forbidden region transfers found in data-flow audits.

B:

When auditors need to see when governance controls were enforced for decisions, the control layer can keep traceable control evidence, access decisions and versioned approvals; however, that proxy would not demonstrate forbidden region transfers found in data-flow audits for the selected intervention.

D:

When EU customer profiling may process identifiable personal data, the control layer can involve privacy counsel in lawful basis, purpose and data minimization assessment; by contrast, the contract administration release needs evidence from forbidden region transfers found in data-flow audits.

Question 906

The defining problem for a supply-chain planning pilot is that large context traffic causes budget spikes without obvious request-count increases. Which proposed approach addresses it most precisely?

  1. Memory separation by trust level
  2. Source-priority truncation policy
  3. Token-economics telemetry
  4. Exact token budget accounting

Explanation

Correct Answer

C:

The mechanism is specific: the model configuration can record uncached input, cache reads, output and tool-related usage per task. A credible implementation requires request-level usage accounting and trace identifiers, while its acceptance test monitors tokens and dollars per accepted outcome by category. The alternative approaches leave the risk that request totals alone cannot isolate token-heavy work.

Incorrect Answers

A:

When user preferences and external tool outputs enter the same conversation memory, the model configuration can store verified state separately from untrusted retrieved text and time-bound hints; the missing operational behavior is specific: the model configuration can record uncached input, cache reads, output and tool-related usage per task.

B:

When a long research input must retain governing rules and current evidence, the model configuration can truncate lowest-priority material while protecting essential instructions and citations; the current decision instead depends on a specific capability: the model configuration can record uncached input, cache reads, output and tool-related usage per task.

D:

If templates unexpectedly exceed request limits after tool results are appended, the model configuration can estimate input, tool-result and output allowances before sending requests; the missing operational behavior is specific: the model configuration can record uncached input, cache reads, output and tool-related usage per task.

Question 907

An education services pilot cannot be accepted because guessing from one screenshot hides true upstream causes. The owner can change one primary architectural control; what should it be?

  1. Introduce Post-incident regression capture
  2. Introduce Targeted trace inspection
  3. Introduce Reproduction-first debugging
  4. Introduce Safe operational rollback

Explanation

Correct Answer

C:

The problem lies in the absence of a mechanism where the engineering workflow can capture minimal reproducible requests and traces before changing behavior. Reinstating it, backed by redacted traces and environment version details, limits the exposure that guessing from one screenshot hides true upstream causes; the engineering team should track reproducible incident rate and time to confirmed root cause.

Incorrect Answers

A:

When a one-off patch without tests allows the same bug to return, the engineering workflow can turn the root cause into a permanent automated test and review the rollout path; yet it would not correct the observed weakness: guessing from one screenshot hides true upstream causes.

B:

This addresses the failure where top-level request durations cannot identify the responsible dependency, because the engineering workflow can correlate model, orchestration and external calls on the critical path; but this remedy operates on another boundary, not the case where guessing from one screenshot hides true upstream causes.

D:

This addresses the failure where debugging in place keeps all users exposed to the broken state, because the engineering workflow can restore a known-good configuration before deeper root-cause analysis; yet it would not correct the observed weakness: guessing from one screenshot hides true upstream causes.

Question 908

A design authority for enterprise knowledge management has established this requirement: many Claude clients need access to the same governed business capability. Which intervention should be prioritized?

  1. Adopt CLI invocation in controlled automation
  2. Adopt MCP server as standardized tool boundary
  3. Adopt Webhooks for event-driven triggers
  4. Adopt Direct Messages API service integration

Explanation

Correct Answer

B:

The proposed control fits because the integration can expose reusable typed tools and resources through a controlled MCP server. Its production prerequisites include server ownership, tool schemas and access control; check authorized tool compatibility and change success across clients on representative tasks. Without this boundary, duplicating bespoke connectors across clients multiplies maintenance.

Incorrect Answers

A:

When a development workflow needs a reproducible one-off command in CI, the integration can use a noninteractive CLI process with bounded permissions and captured exit status; whereas the enterprise knowledge management case must guard against the outcome that duplicating bespoke connectors across clients multiplies maintenance.

C:

When new repository events should start work without constant polling, the integration can accept validated inbound events and enqueue bounded downstream tasks; it does not ensure the required behavior: the integration can expose reusable typed tools and resources through a controlled MCP server.

D:

When a backend requires tightly scoped Claude inference inside its existing request pipeline, the integration can call the model through the application service with controlled retries and validation; whereas the enterprise knowledge management case must guard against the outcome that duplicating bespoke connectors across clients multiplies maintenance.

Question 909

A design authority for internal audit has established this requirement: typical requests succeed but rare regulatory exceptions cause serious errors. Which intervention should be prioritized?

  1. Adopt Reconciliation of conflicting sources
  2. Adopt Exception-first task splitting
  3. Adopt Map-reduce summarization
  4. Adopt Dependency graph scheduling

Explanation

Correct Answer

B:

The intervention acts on the system behavior directly: the architecture can separate normal and exception branches with stronger checks for the latter. Before rollout, verify an exception taxonomy and differentiated routing policy. The relevant success test is exception detection recall and unsafe bypass counts; otherwise an average-case prompt can hide high-impact failure cases.

Incorrect Answers

A:

When two authoritative repositories return different effective dates, the architecture can separate conflict detection and resolution from answer drafting; the stated scenario depends on a different mechanism: the architecture can separate normal and exception branches with stronger checks for the latter.

C:

If thousands of independent notes exceed practical context budgets, the architecture can summarize bounded chunks then aggregate their structured findings; whereas the energy operations case must guard against the outcome that an average-case prompt can hide high-impact failure cases.

D:

If downstream analyses sometimes run before upstream data cleansing finishes, the architecture can model the task as a DAG and execute nodes only when prerequisites are satisfied; the stated scenario depends on a different mechanism: the architecture can separate normal and exception branches with stronger checks for the latter.

Question 910

During technical sign-off for healthcare operations, Parallelizable workflow optimization has been selected. The design premise is that the evaluation process can parallelize work while joining outputs under a bounded concurrency cap. Which enabling condition must be demonstrated before production traffic is allowed?

  1. Confirm durable reconciliation and batch failure handling
  2. Confirm relevance evaluation and source-preserving deduplication
  3. Confirm independence proof and capacity testing
  4. Confirm traceable retry and human-review expense

Explanation

Correct Answer

C:

The prerequisite creates the necessary operating boundary: independence proof and capacity testing. The system is intended to work because the evaluation process can parallelize work while joining outputs under a bounded concurrency cap. A release without that foundation may still suffer blindly increasing concurrency can overload downstream quotas; use overall task latency and quota error rate to evaluate the result.

Incorrect Answers

A:

Where offline processing runs thousands of independent requests nightly, the evaluation process can group asynchronous requests using the available batch workflow; however, the named intervention needs independence proof and capacity testing before its protection can be enforced.

B:

For the separate challenge where redundant retrieved passages consume most of the prompt budget, the evaluation process can deduplicate and rank passages before model invocation; but the production dependency for Parallelizable workflow optimization is independence proof and capacity testing.

D:

The setup is useful to enable the alternative's behavior where the evaluation process can compare token savings against full correction and review cost and shaving tokens degrades downstream extraction precision; however, the named intervention needs independence proof and capacity testing before its protection can be enforced.

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