Anthropic CCA-F (Claude Certified Architect - Foundations (CCAR-F)) Practice Test 8

Question 70

A workflow can validate a shipping address locally, then purchase a nonrefundable label. Which ordering reduces avoidable irreversible mistakes?

  1. Complete validation and required approval before purchasing the label.
  2. Purchase the label and validate only if the carrier rejects it.
  3. Run validation and purchase concurrently.
  4. Let the model choose an arbitrary order on each run.

Explanation

Correct Answer

A:

Deterministic preconditions should precede a costly commitment. The execution boundary should also reject stale or unapproved inputs.

Incorrect Answers

B:

A carrier may accept a syntactically valid but incorrect address, leaving an unnecessary purchase.

C:

The purchase depends on the validation result; concurrency removes that protection.

D:

A known prerequisite should not be left to variable planning.

Question 71

An agent can propose, approve, execute, and verify a change. Its controller allows execution directly from proposed state. What is the architectural defect?

  1. The workflow uses separate proposal and approval records instead of one record.
  2. Every state must use a separate model.
  3. Verification must occur before proposal.
  4. A required state transition can be bypassed.

Explanation

Correct Answer

D:

The controller must enforce allowed transitions and their evidence. Listing stages in a prompt does not ensure the approval state was reached.

Incorrect Answers

A:

Separate records can support a valid state machine; the defect is allowing execution without the required approved transition.

B:

State enforcement can be deterministic and does not require one model per stage.

C:

Verification of an executed change requires the relevant operation and resulting state.

Question 72

A plan is approved, but the user later changes the target account. What should happen before executing account-specific actions?

  1. Use whichever account appeared first in the conversation.
  2. Execute against both accounts to preserve both interpretations.
  3. Reuse approval because the general action type is unchanged.
  4. Revalidate the changed scope and obtain any newly required approval.

Explanation

Correct Answer

D:

Approval must match the actual action and target. A changed account can alter permissions, consequences, and the user's intended commitment.

Incorrect Answers

A:

Historical order does not resolve the user's current instruction.

B:

This expands scope beyond the request and introduces an unintended second action.

C:

Authorization for one target does not necessarily cover another target.

Question 73

A research-then-publish workflow exposes publishing tools during exploratory research. What design reduces premature publication risk?

  1. Provide more publication examples during research.
  2. Require the model to mention research in every response.
  3. Expose publication capability only after the required review state is reached.
  4. Leave the publish tool callable but suppress its result from the research worker's transcript.

Explanation

Correct Answer

C:

Phase-specific capabilities narrow what can execute during exploration. The publication service should still enforce its own authorization and preconditions.

Incorrect Answers

A:

Examples can teach syntax but do not enforce the workflow boundary.

B:

Response wording does not remove executable publishing capabilities.

D:

Hiding the observation does not prevent the external publication and makes an accidental action harder to reconcile.

Question 74

An agent must calculate invoice tax from specified rates and rounded line amounts. Which component should provide the authoritative arithmetic?

  1. A majority vote among unverified generated totals.
  2. A tested deterministic calculation function using the declared rounding rules.
  3. The longest available context window.
  4. The most fluent generated numerical explanation.

Explanation

Correct Answer

B:

The model can choose the function and explain the result, but reproducible arithmetic should follow explicit rules in code.

Incorrect Answers

A:

Agreement can repeat the same arithmetic error and does not enforce the accounting rule.

C:

More context does not itself perform or verify deterministic calculations.

D:

Fluency does not establish numerical correctness or consistent rounding.

Question 75

A workflow must classify free-text requests and enforce a fixed spending ceiling. Which division of responsibility is strongest?

  1. Ask the model to perform both tasks without backend checks.
  2. Use code to guess intent and let the model decide whether limits apply.
  3. Use model reasoning for classification and deterministic enforcement for the ceiling.
  4. Reject all free-text requests before classification.

Explanation

Correct Answer

C:

Probabilistic interpretation suits ambiguous language; a fixed financial boundary belongs in code that validates the actual transaction.

Incorrect Answers

A:

Language understanding can help classification, but it cannot substitute for enforceable spending policy.

B:

The spending rule is deterministic and should not depend on discretionary interpretation.

D:

This avoids the application function rather than implementing the two required controls.

Question 76

A summarization task needs no external actions and already meets quality targets in one request. What is the best next architecture step?

  1. Add a coordinator and specialist for every paragraph before measuring whether the single-request baseline has any material failure.
  2. Keep the simple baseline and add complexity only for demonstrated shortcomings.
  3. Force at least one tool call so the system qualifies as agentic.
  4. Replace the evaluated prompt with an unevaluated autonomous loop solely to reduce the amount of application code.
  5. Remove evaluation because the workflow is simple.

Explanation

Correct Answer

B:

Architecture should be justified by measured requirements. A successful single-call design is a useful baseline for later comparisons.

Incorrect Answers

A:

The extra handoffs add cost and complexity without evidence that the existing architecture is insufficient for the task.

C:

A label does not create user value; unnecessary tool use adds work.

D:

Less application code does not establish a better result or lower total complexity. The existing measured baseline should be displaced only for a demonstrated benefit.

E:

Simple systems still need monitoring and regression checks.

Question 77

A router cannot confidently classify a request between billing and account recovery. Neither workflow safely handles the other category. Which design is appropriate?

  1. Choose billing whenever confidence values tie.
  2. Increase the router's response length without changing its decision policy.
  3. Run both workflows with full write access.
  4. Use a clarification or safe fallback path before specialist execution.

Explanation

Correct Answer

D:

Ambiguous classification should not silently trigger consequential actions. A bounded fallback can resolve intent or escalate uncertainty.

Incorrect Answers

A:

An arbitrary tie-break can route sensitive work to an unsuitable procedure.

B:

More prose does not establish which category is correct.

C:

Parallel execution can produce conflicting or unauthorized account actions.

Question 78

A router returns a category outside the application's supported set. What should the controller do?

  1. Validate the category and route invalid values through a defined recovery path.
  2. Create a new production workflow from the returned text automatically.
  3. Map every unsupported value to the default workflow without checking its suitability.
  4. Treat the returned label as a shell command.

Explanation

Correct Answer

A:

The controller should not execute an undefined branch. A schema or allowlist and explicit fallback make invalid routing observable and recoverable.

Incorrect Answers

B:

An unrecognized label does not authorize a new execution path.

C:

A silent default can execute an inappropriate procedure for an invalid routing result.

D:

A classification value is data, not executable control text.

Question 79

A planner says task A requires B's result, B requires C's result, and C requires A's result. None has an initial result. What should the scheduler do?

  1. Assign each task more model tokens while keeping the same prerequisites.
  2. Detect the dependency cycle and revise the plan to supply or remove a justified prerequisite.
  3. Start task A while treating its missing prerequisite as an empty successful result.
  4. Launch all three and wait without a deadline.

Explanation

Correct Answer

B:

A cyclic dependency graph needs a deliberate break or reformulation before executable scheduling is possible.

Incorrect Answers

A:

Additional generation capacity cannot create an unavailable predecessor result.

C:

Fabricating a predecessor result bypasses the dependency rather than making the plan executable.

D:

No task can obtain its prerequisite, so launching the cycle does not produce progress.

Question 80

A dispatcher acknowledges a submitted analysis job before recording it anywhere durable. Its process then crashes, and the job disappears. Which ordering prevents this loss?

  1. Mark the job complete when acknowledgment is sent.
  2. Persist an accepted job in a recoverable queue or store before acknowledging acceptance.
  3. Acknowledge earlier and rely on the caller never checking again.
  4. Keep the job only in a model's conversation summary.

Explanation

Correct Answer

B:

Durable acceptance lets another process recover the job after a dispatcher failure.

Incorrect Answers

A:

Acceptance records an obligation to process work, not evidence that processing finished.

C:

An earlier acknowledgment widens the period in which accepted work can vanish.

D:

A generated summary is not a durable dispatch record with recoverable execution semantics.

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