Anthropic CCA-F (Claude Certified Architect - Foundations (CCAR-F)) Practice Test 12

Question 110

A worker returns malformed JSON where a downstream calculator expects numeric fields. What should happen before calculation?

  1. Extract every number with a regular expression and assign fields by appearance order.
  2. Pass the raw text and hope the calculator infers the intended values.
  3. Validate and repair or reject the worker output against the agreed contract.
  4. Replace every missing number with zero without a documented rule.

Explanation

Correct Answer

C:

Downstream computation needs well-defined inputs. A schema gate prevents malformed data from becoming an apparently precise numerical result.

Incorrect Answers

A:

Numeric appearance does not establish which contract field each value belongs to or whether required fields are present.

B:

Implicit coercion can conceal missing or misidentified fields.

D:

Zero is a substantive value and may distort the calculation.

Question 111

An agent reaches its configured resource ceiling with two subtasks unfinished. What should its final status convey?

  1. Completed work, unfinished scope, and the reason execution stopped.
  2. Only the number of tool calls made.
  3. Successful completion because the budget was fully used.
  4. No status at all so the user cannot see the limit.

Explanation

Correct Answer

A:

A clear bounded outcome lets the user assess usefulness and next steps. The system should not invent completion to satisfy a success label.

Incorrect Answers

B:

Activity counts omit the task outcome and remaining work.

C:

Spending the allocation does not establish that the requested work finished.

D:

Silence obscures what was accomplished and what remains.

Question 112

An assistant asks a user to approve deleting a record named Project Alpha, but two records share that name. What should precede approval?

  1. Ask for approval again using the same ambiguous name.
  2. Choose the first search result silently.
  3. Resolve the exact record and present identifying details for the intended target.
  4. Request approval of the shared display name and defer exact identity selection until execution.

Explanation

Correct Answer

C:

Approval must refer to a concrete action on a distinguishable object. An ambiguous label cannot establish which record the user authorized deleting.

Incorrect Answers

A:

Repeating the question does not resolve the missing identity information.

B:

Ranking does not prove the user's intended target.

D:

Approval should concern an identifiable action; selecting among different records afterward can change what the user actually approved.

Question 113

A team evaluates an agent on recorded purchase requests. How should it avoid placing real orders during evaluation?

  1. Use production purchasing credentials and cancel orders afterward.
  2. Remove all tool checks from the test harness.
  3. Trust the model to recognize that the run is a test.
  4. Use controlled tool substitutes or sandbox endpoints with explicit nonproduction credentials.

Explanation

Correct Answer

D:

The harness can exercise decision paths without real commitments. Test substitutes should model relevant success and failure behavior accurately.

Incorrect Answers

A:

Cancellation may be costly, unavailable, or incomplete; evaluation should not depend on undoing real purchases.

B:

This prevents meaningful evaluation of the action path.

C:

A prompt cue is weaker than controlling executable capabilities.

Question 114

An agent passes evaluation with a mock inventory tool that always returns unlimited stock. Production frequently returns partial stock. What should the harness change?

  1. Keep the mock unchanged because deterministic success makes comparisons easier.
  2. Grade the agent's politeness instead of its inventory decisions.
  3. Model relevant production outcomes, including partial stock, with explicit expected actions.
  4. Only add longer successful inventory descriptions.

Explanation

Correct Answer

C:

Representative tool semantics exercise decisions the agent will actually face after deployment.

Incorrect Answers

A:

A convenient mock can conceal the production branch that determines whether the agent behaves correctly.

B:

Presentation quality cannot validate the absent business behavior.

D:

More wording does not introduce the missing partial-availability condition.

Question 115

A workflow routes requests to either read-only reporting or account maintenance. What should happen to tool availability after routing?

  1. Remove all tools from both branches.
  2. Use tool ordering as the only restriction.
  3. Provide each branch only the capabilities needed for its authorized function.
  4. Expose every tool to both branches for convenience.

Explanation

Correct Answer

C:

Routing can narrow the execution surface as well as the prompt. Backend permissions remain necessary even when tool exposure is restricted.

Incorrect Answers

A:

The maintenance branch may require authorized operations to fulfill its task.

B:

Placement in a list does not prevent invoking a capability.

D:

This grants unnecessary capabilities to the reporting branch.

Question 116

An agent trace needs enough detail for debugging but includes customer secrets. What logging design is most appropriate?

  1. Let the model decide ad hoc which secrets are safe to publish.
  2. Disable all logging, including operation IDs and failures.
  3. Record necessary event metadata and redact or protect sensitive payloads under a defined policy.
  4. Log every raw payload indefinitely.

Explanation

Correct Answer

C:

Observability should preserve diagnostic value while limiting exposure. Access controls and retention rules belong alongside redaction.

Incorrect Answers

A:

Sensitive logging needs consistent policy and enforcement rather than discretionary generation.

B:

This removes useful accountability that can often be retained without secret content.

D:

Unbounded sensitive retention increases privacy and security risk without necessarily improving debugging.

Question 117

A payment failure triggers a compensating reservation release, but the release itself fails. What state should the workflow record?

  1. The entire workflow succeeded because both forward and reverse steps were invoked.
  2. An unresolved compensation with its operation identity and a supported retry or manual-repair path.
  3. The original reservation never existed because compensation was attempted.
  4. A new unrelated reservation to balance the count.

Explanation

Correct Answer

B:

Failed compensation requires explicit follow-up rather than being hidden as a completed rollback.

Incorrect Answers

A:

Invocation of unsuccessful steps does not establish the intended final business state.

C:

Attempting a reversal does not erase the already committed reservation.

D:

Creating another resource does not repair the failed release and adds further side effects.

Question 118

A write endpoint confirms an update, but a search index still shows the old value immediately afterward. What is the best interpretation when indexing is asynchronous?

  1. Repeat the write with a different target until search changes.
  2. Ignore every future search result permanently.
  3. Treat any immediate search mismatch as proof that the write endpoint's success response was fabricated, without inspecting the authoritative record.
  4. Use authoritative readback or operation status and account for the documented indexing delay.
  5. Treat the search index as strongly consistent even though its documented update path is asynchronous.

Explanation

Correct Answer

D:

Verification should query the right consistency boundary. A stale secondary view does not by itself negate a confirmed commit.

Incorrect Answers

A:

This can introduce duplicate or incorrect modifications.

B:

The index remains useful once its freshness constraints are understood.

C:

The stated indexing delay already explains a temporary mismatch. Consistency behavior must be distinguished from evidence of a failed write.

E:

This assumption applies a guarantee the index does not provide. An immediate stale result can reflect indexing lag rather than a failed authoritative write.

Question 119

An agent proposes a high-cost fan-out that would exceed the remaining task budget. Where should the limit be enforced?

  1. Only in a final cost summary after execution.
  2. At scheduling or execution before launching work that exceeds the allocation.
  3. Give each parallel worker the entire remaining budget and total usage afterward.
  4. Reserve the full allocation independently for every child job, without subtracting other pending reservations.

Explanation

Correct Answer

B:

Budget checks must precede resource commitment. The coordinator can reduce scope, choose a cheaper path, or report the limit.

Incorrect Answers

A:

Post-hoc reporting cannot prevent spending beyond the allowed ceiling.

C:

Independent full-budget grants can collectively exceed the task-wide ceiling before reconciliation.

D:

Concurrent reservations can collectively exceed the remaining budget even before spending occurs. The scheduling decision must account for work already admitted.

Question 120

A long-running task resumes after its workflow definition changes. Why should the checkpoint record the workflow version?

  1. To ensure the model never receives new instructions.
  2. To interpret saved state under the correct transition and data contracts.
  3. To make every previous result automatically invalid.
  4. To replace all business identifiers with version numbers.

Explanation

Correct Answer

B:

A checkpoint created by one workflow may not map safely to another. Version-aware migration or resumption avoids applying new assumptions to old state silently.

Incorrect Answers

A:

Version tracking does not prohibit deliberate updates; it makes them controlled.

C:

Some results remain valid; compatibility must be assessed rather than assumed absent.

D:

Workflow identity and business-object identity serve different purposes.

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