Anthropic CCA-F (Claude Certified Architect - Foundations (CCAR-F)) Practice Test 3

Question 20

An extraction prompt requests 'all details' from invoices, producing irrelevant fields. What redesign reduces drift?

  1. Describe evaluation priorities such as latency, correctness, budget and operational ownership
  2. Enumerate the exact fields, types and normalization rules required downstream
  3. Mandate verbatim preservation for named entities while normalizing only designated fields
  4. Require source-backed claims and a labeled unknowns or uncertainty section
  5. Specify redaction rules and permitted data fields in the expected summary

Explanation

Correct Answer

B:

Explicit extraction targets guide attention to operationally meaningful values. A vague request leaves the output schema and business semantics undefined.

Incorrect Answers

A:

Architectural choices depend on goals and trade-offs; supporting making architecture trade-off priorities explicit cannot replace the required defining bounded extraction fields.

C:

Field-level instructions distinguish transcription from interpretation; that serves preserving exact source spelling where required, whereas the case requires defining bounded extraction fields.

D:

Because requiring evidence mapping and explicit uncertainty discourages treating unverified inference as fact, the measure supports separating sourced conclusions from unknowns rather than the needed defining bounded extraction fields.

E:

Data minimization instructions make the output contract sensitive to privacy; its benefits for minimizing sensitive data in summaries do not resolve the need for defining bounded extraction fields.

Question 21

An organization wants to measure whether the agent's low-confidence escalation threshold works. What data should it collect?

  1. Choose automatic acceptance for verified low-risk cases and exception-based human review
  2. Review samples by document type and route high-risk unfamiliar cases for deeper inspection
  3. Display policy basis, supporting records, and the requested action before approval
  4. Compare escalation outcomes with labeled ground truth and false-positive rates

Explanation

Correct Answer

D:

Thresholds should be calibrated empirically, not chosen solely from model self-descriptions. Outcome tracking reveals unnecessary escalations and missed risky cases.

Incorrect Answers

A:

Hybrid workflows reserve human effort for uncertain or high-impact outcomes; that may serve balancing automation with targeted human oversight, but it is no substitute for calibrating escalation thresholds against observed outcomes.

B:

Aggregate accuracy can hide weak performance on small important cohorts; this may support targeting review to difficult high-impact document cohorts but cannot alone meet the requirement for calibrating escalation thresholds against observed outcomes.

C:

Meaningful approval requires adequate context to judge eligibility and consequences; the stated problem requires calibrating escalation thresholds against observed outcomes rather than giving human reviewers sufficient decision evidence.

Question 22

A Claude Code safety gate must reject dangerous Bash invocations before they run while leaving unrelated tool calls unaffected. Which TWO configuration decisions achieve both controls? Choose TWO.

  1. Match the safety hook to Bash calls (or the specific dangerous command pattern), not every tool
  2. Check protected file paths only when a filesystem tool runs, without examining unsafe Bash commands
  3. Reject matching dangerous shell requests through a blocking PreToolUse hook
  4. Log Bash outcomes after the tool has completed and use the log as the safety gate

Explanation

Correct Answers

A:

The design must include targeting a hook to relevant operations: matchers determine which hook executions apply; tight scoping prevents policy handlers from disrupting operations they were not designed to supervise.

C:

PreToolUse runs before a tool call and can participate in the decision to block it. A post-execution observer cannot reliably prevent the command's side effects.

Incorrect Answers

B:

A path gate can protect some resources, but the stated requirement is to reject dangerous Bash invocations before execution. This choice does not cover dangerous shell commands in general.

D:

Post-execution monitoring cannot prevent an unsafe command from running. The needed block belongs at the pre-execution hook point.

Question 23

Two documents have the same title but different revision dates, and an agent quotes the older policy. What evidence handling prevents this?

  1. Caveat the claim and request stronger corroboration before treating it as established
  2. Record only the time the web page was visited, ignoring its policy revision and effective date
  3. Mandate a matching source span or flag the clause as unsupported
  4. Interpret retrieved document instructions as untrusted content rather than authority
  5. Track source version and effective date for policy-dependent statements

Explanation

Correct Answer

E:

Correct citations must identify the applicable revision, not just a title. Version metadata helps reviewers detect obsolete guidance before it affects decisions.

Incorrect Answers

A:

Weak or incomplete evidence warrants explicit uncertainty; the unsatisfied requirement is maintaining revision provenance for policy evidence, not calibrating claims to evidence strength.

B:

A retrieval timestamp alone does not identify which policy version governed the case. The controlling revision and effective date must be checked before a policy-dependent claim is made.

C:

When considering tracing extracted clauses to document spans, evidence-grounded extraction ties fields to document locations; the requirement for maintaining revision provenance for policy evidence remains unmet.

D:

When addressing resisting prompt injection embedded in source material, external evidence may contain prompt injection; the present task instead calls for maintaining revision provenance for policy evidence.

Question 24

The team wants CI to block merges only for reproducible, high-severity defects, not style suggestions. What prompt design helps?

  1. Define explicit severity criteria and require evidence for blocking findings
  2. Choose least-privilege credentials and sandbox the untrusted review workload
  3. Mask secrets and keep tokens in the runner's protected secret store
  4. Insist on file path, line range and finding rationale in a validated result schema

Explanation

Correct Answer

A:

Reviewer outputs should distinguish verified blockers from optional recommendations. Clear rubrics and supporting file references reduce false-positive merge failures.

Incorrect Answers

B:

PR diffs can contain prompt injection or malicious scripts; limiting trust for untrusted pull-request code and calibrating automated code-review severity are distinct responsibilities under these constraints.

C:

Build logs are often retained and widely visible; the action supports keeping CI credentials out of generated logs but falls short of calibrating automated code-review severity in this case.

D:

Location-aware structured findings allow a bot to attach actionable comments; this can serve generating location-specific PR findings, but leaves the requirement for calibrating automated code-review severity unmet.

Question 25

An agent needs to distinguish waiting for a user from waiting for a tool. Which state design is appropriate?

  1. Send the case to an explicit policy-resolution or human-review state
  2. Treat eligibility verification as a required prerequisite in the workflow graph
  3. Mandate a citation audit to pass before publishing the report
  4. Represent external-input and tool-pending states separately

Explanation

Correct Answer

D:

Different waits have different timeouts, retries and authorization requirements. A single generic waiting state obscures when it is safe to continue autonomously.

Incorrect Answers

A:

Ambiguity requires a governed branch instead of a silent best guess; that serves modeling ambiguity as an escalation branch, whereas the case requires distinguishing human and tool handoff states.

B:

Because an explicit dependency ensures the mutation cannot begin before policy checks succeed, the measure supports enforcing refund prerequisites before mutation rather than the needed distinguishing human and tool handoff states.

C:

A publish gate makes source validity an explicit dependency; its benefits for gating publication on source validation do not resolve the need for distinguishing human and tool handoff states.

Question 26

A payment status lookup times out, but the request might have reached the service. Which retry design avoids duplicate financial actions?

  1. Combine bounded retries with an operation idempotency key and outcome lookup
  2. Provide an explicit failed tool result with diagnostic information rather than fabricate data
  3. Retain the tool implementation in the authorized execution environment
  4. Require a step budget and a defined fallback when the loop exceeds it

Explanation

Correct Answer

A:

Retries of side-effecting actions can duplicate work when acknowledgment is lost. Idempotency and reconciliation separate genuine failures from already-completed operations before repeating them.

Incorrect Answers

B:

The tool-error signal should flow into the conversation so the agent can retry, escalate, or communicate limits; a method suited to surfacing client-tool execution failure does not meet the need for safe recovery after an ambiguous tool timeout.

C:

For distinguishing model intent from tool execution, client tools are interfaces advertised to the model; the hosting application runs their operations; however, this requirement concerns safe recovery after an ambiguous tool timeout.

D:

Independent limits on turns, time, and side effects prevent runaway agent cycles; that may serve bounding an autonomous agent loop, but it is no substitute for safe recovery after an ambiguous tool timeout.

Question 27

A receipt parser returns line items whose sum differs from the reported invoice total. What validation should run?

  1. Reissue with targeted feedback identifying the invalid field and expected format
  2. Ask for a better source or human review instead of repeated identical retries
  3. Recalculate totals and reject or review arithmetic inconsistencies
  4. Categorize failure types and retry only plausibly transient conditions

Explanation

Correct Answer

C:

Semantic checks compare extracted values to business invariants. A schema can confirm numeric types but cannot prove the figures reconcile.

Incorrect Answers

A:

Although feedback-based repair gives the model a specific correctable defect, the choice addresses targeting a retry at a correctable parse defect and not the requested checking arithmetic consistency in extracted records.

B:

An absent source page cannot be recovered by sampling the same incomplete input; the deciding condition is checking arithmetic consistency in extracted records, whereas this action concerns recognizing a nonrecoverable extraction input gap.

D:

Authentication and permission errors require configuration or credential repair; this addresses distinguishing retryable and nonretryable failures, while the question is about checking arithmetic consistency in extracted records.

Question 28

One research subtask fails repeatedly, but the others finished. Which scheduler policy minimizes wasted work?

  1. Bundle closely related low-cost tasks into a meaningful worker assignment
  2. Retry or replace only the failed subtask while keeping successful outputs
  3. Split read-only verification from the authorized state-changing step
  4. Order exploration, plan approval and implementation as dependent tasks

Explanation

Correct Answer

B:

Independence in the task graph permits selective replay. Restarting the whole research pipeline multiplies cost and may change valid evidence unnecessarily.

Incorrect Answers

A:

When considering avoiding overfragmentation of delegated work, excessive decomposition multiplies setup and context cost; the requirement for isolating retry to failed work remains unmet.

C:

Distinct stages expose eligibility decisions and permit human approval before mutations; separating validation from side-effecting tasks and isolating retry to failed work require different controls in this situation.

D:

Staging the work lets decisions use verified repository context; the unsatisfied requirement is isolating retry to failed work, not decomposing dependent coding stages.

Question 29

A multi-agent report loses the customer's locale when tasks are delegated. How should the coordinator repair the contracts?

  1. Include locale and other required business constraints in each relevant worker brief
  2. Mandate evidence of which documents and versions the worker actually consulted
  3. Choose a specialized subagent with only the necessary non-executing inspection tools
  4. Provide condensed relevant findings and source handles rather than full unrelated transcripts

Explanation

Correct Answer

A:

Constraints do not reliably propagate by implication across context boundaries. Explicit parameters let every delegated answer apply the same customer-specific policy.

Incorrect Answers

B:

Source and version metadata distinguish inspected facts from memory-based guesses; this can serve validating the evidentiary basis of a subagent answer, but leaves the requirement for propagating shared constraints into tasks unmet.

C:

Tool access should match the worker's responsibility; that addresses preventing command execution in a read-only worker, whereas this case calls for propagating shared constraints into tasks.

D:

Selecting task-relevant evidence controls token growth without erasing provenance; this concerns controlling delegated context size, leaving the requirement for propagating shared constraints into tasks outstanding.

Question 30

A coordinator has three worker results but no rule for selecting the final recommendation. What is the missing architectural component?

  1. Partition independent expert analyses and reserve integration for the orchestrator
  2. Send trivial deterministic steps directly and reserve specialists for nontrivial reasoning
  3. Add a deterministic aggregation policy for worker output acceptance
  4. Focus on holding publication until validation reports are reconciled with the extracted artifacts

Explanation

Correct Answer

C:

A merge contract can specify completeness, conflict handling and evidence precedence. Without it, the final recommendation may depend on incidental completion order.

Incorrect Answers

A:

Focused workers can receive targeted instructions and smaller context windows; its benefits for matching specialist agents to distinct competencies do not resolve the need for defining a reproducible worker merge policy.

B:

Agent delegation has token, latency and observability costs; supporting avoiding needless specialist agent invocations cannot replace the required defining a reproducible worker merge policy.

D:

Completion of one branch does not imply total workflow completion; that serves awaiting validation across distributed workers, whereas the case requires defining a reproducible worker merge policy.

Claude Certified Architect - Foundations (CCAR-F) (CCA-F) Practice Tests