Anthropic CCA-F (Claude Certified Architect - Foundations (CCAR-F)) Practice Test 6

Question 50

An extraction assistant assigns confidence 0.95 to many records, but human review finds that only about 70% of those records are correct. The team wants to use confidence to route risky records for review. What should it do first?

  1. Reduce review volume by accepting every record above the current 0.90 threshold.
  2. Interpret 0.95 as a guaranteed 95% correctness probability for each record.
  3. Measure score reliability on representative labeled cases and choose review thresholds from observed error rates.
  4. Average several confidence statements from the same generation and use their mean as ground truth.

Explanation

Correct Answer

C:

Calibration relates reported confidence to actual outcomes. Thresholds should then reflect the cost of errors and review capacity, with continued monitoring for distribution changes.

Incorrect Answers

A:

The current scores overstate reliability in the described group. Applying a lower unvalidated cutoff would not establish an acceptable risk level.

B:

The observed error rate contradicts that interpretation. A model-reported score is not automatically calibrated to empirical correctness.

D:

Repeated self-assessments can share the same bias. Averaging them does not provide the labeled outcomes needed to measure reliability.

Question 51

A custom agent loop receives stop_reason end_turn with a final answer and no pending tool calls. What should the controller normally do?

  1. Treat the response as an API failure.
  2. Send an empty tool_result message.
  3. Return the answer and finish the turn.
  4. Repeat the identical request until it requests a tool.

Explanation

Correct Answer

C:

The model has completed this turn. Further inference needs a new task or an explicit application-level reason to continue.

Incorrect Answers

A:

Normal completion is a successful outcome, distinct from transport errors and interrupted generation.

B:

No invocation is waiting for an observation. An empty result does not represent an executed operation.

D:

A task can finish without tools. Requiring tool use would manufacture unnecessary work after completion.

Question 52

A Claude Messages response has stop_reason tool_use and a content array containing introductory text followed by two client tool_use blocks. Both requested operations are permitted. The controller reads only content[0] and returns the text as its final answer. What fixes this defect?

  1. Return the text and wait for a new user message before examining the remaining blocks.
  2. Inspect the typed content blocks and dispatch both tool_use requests through the normal tool loop.
  3. Process only the last content block because it is the most recent instruction.
  4. Parse the introductory sentence for tool names and reconstruct the requested arguments.

Explanation

Correct Answer

B:

Introductory text can coexist with pending tool requests. Reading the full content array prevents the controller from mistaking narration for completion and silently dropping the requested operations.

Incorrect Answers

A:

The response already requests tool execution. Waiting for an unrelated user turn leaves the current tool loop incomplete.

C:

Array position does not make an earlier tool_use block obsolete. This would leave one of the two invocations unanswered.

D:

The typed blocks already carry the invocation details. Reconstructing calls from prose can lose parameters or dispatch an operation that the prose merely describes.

Question 53

A supported server-tool workflow returns stop_reason pause_turn. No client tool call is waiting for the application. What should the controller do if work remains within budget?

  1. Continue using the returned assistant content through the documented continuation flow.
  2. Discard the returned content and always restart the original task from scratch.
  3. Fabricate a client tool_result for the unfinished server-side operation.
  4. Treat pause_turn as equivalent to a final end_turn response.

Explanation

Correct Answer

A:

Preserving the paused response lets the server-side workflow resume without pretending it completed.

Incorrect Answers

B:

A restart loses the continuation state and can repeat work already performed.

C:

A server-managed pause is not a request for the application to invent a client-tool observation.

D:

A paused server-tool turn does not establish that the answer is finished.

Question 54

After executing a requested tool, a controller sends its result but omits the assistant message containing the tool_use block. Why can this fail?

  1. The result should use a newly generated invocation ID rather than the ID returned by the model.
  2. The tool result must repeat every tool definition verbatim.
  3. The result lacks its corresponding invocation in the conversation history.
  4. Tool results must always be sent as system messages.

Explanation

Correct Answer

C:

The controller must preserve the assistant's tool request before returning its correlated observation. Otherwise the result references a call the conversation does not contain.

Incorrect Answers

A:

A new identifier breaks correlation with the invocation. It cannot compensate for removing the message that introduced the original tool-use ID.

B:

Definitions and invocation history serve different purposes. Repetition does not restore the missing call.

D:

Client-tool observations belong in user messages, not privileged instructions.

Question 55

A workflow must look up a customer by email before querying orders by the returned customer ID. No ID is initially available. Which schedule is valid?

  1. Resolve the customer, validate the result, then query orders with its ID.
  2. Start both calls together using the email as the order query's ID.
  3. Query orders with a guessed identifier and verify afterward.
  4. Run both lookups together, then substitute the resolved ID into the already completed order result if it differs from the guessed one.
  5. Use the most recent customer ID from an unrelated session.

Explanation

Correct Answer

A:

The second operation consumes the first operation's output. Scheduling must preserve this data dependency even if both operations are read-only.

Incorrect Answers

B:

The order endpoint requires the resolved identifier; an email is not a substitute for that value.

C:

Guessing can retrieve the wrong person's records and cannot satisfy the required identity resolution.

D:

Changing a label after execution does not change which customer's orders were actually queried. The dependent call needs the resolved ID before it runs.

E:

Session recency does not establish that the identifier belongs to this customer.

Question 56

Two workers may update the same configuration key. Each starts from the same old version. Which control prevents one from silently overwriting the other's accepted change?

  1. Use different worker names but the same unguarded write endpoint.
  2. Give each worker a longer prompt describing collaboration.
  3. Accept whichever write finishes last.
  4. Require a version precondition and reconcile a rejected conflicting write.

Explanation

Correct Answer

D:

Optimistic concurrency checks reject updates based on stale state. A conflict must be resolved against the new version rather than blindly retried.

Incorrect Answers

A:

Distinct identities improve attribution but do not enforce concurrency control.

B:

Guidance does not make the storage operation detect stale versions.

C:

Last-writer-wins can discard a valid earlier update without recognizing the conflict.

Question 57

Several coding workers will modify independent modules, but all also want to rewrite a shared dependency file. How should the coordinator reduce avoidable merge conflicts?

  1. Let workers overwrite it in completion order.
  2. Ask every worker to regenerate the dependency file independently.
  3. Assign shared-file ownership to one integrator and give workers explicit module boundaries.
  4. Remove the dependency file from review.

Explanation

Correct Answer

C:

Clear ownership limits competing edits. The integrator can reconcile dependency requirements after receiving each module's changes.

Incorrect Answers

A:

Completion order does not establish which combined dependency changes are correct.

B:

Independent rewrites of shared state can conflict even when module edits are otherwise disjoint.

D:

Ignoring the shared artifact hides integration risk rather than preventing it.

Question 58

A report requires evidence from three mandatory sources. Two workers finish; the third times out. Which coordinator response preserves the report's stated scope?

  1. Duplicate one completed worker's result to fill the missing source slot.
  2. Publish the two-source report as fully verified.
  3. Count the timeout as confirmation that the third source contains no relevant evidence.
  4. Retry or escalate the missing work within budget, otherwise label the report incomplete.

Explanation

Correct Answer

D:

The coordinator must distinguish unavailable evidence from completed research. A partial result can still be useful when its coverage limit is explicit.

Incorrect Answers

A:

Reusing evidence does not provide independent coverage of the absent source.

B:

The missing mandatory source leaves a known coverage gap that a completion label would conceal.

C:

A timeout is an execution failure, not a negative search result.

Question 59

Two workers evaluate disjoint ticket samples. Worker A reports 90 correct classifications out of 100; worker B reports 10 out of 20. The coordinator needs the accuracy across all 120 tickets. How should it combine their results?

  1. Average the two reported percentages, giving an overall accuracy of 70%.
  2. Use 90%, because the larger worker sample dominates the combined result.
  3. Use 50%, because the lowest worker accuracy is the conservative combined estimate.
  4. Add the correct counts and sample sizes, giving 100 divided by 120, or about 83.3%.

Explanation

Correct Answer

D:

Each ticket contributes once to the combined numerator and denominator. Returning counts with worker metrics makes aggregation valid when sample sizes differ.

Incorrect Answers

A:

An unweighted mean gives the 20-ticket sample as much influence as the 100-ticket sample. It answers a different question from accuracy across individual tickets.

B:

The larger sample deserves more weight, but the smaller sample still contains 20 evaluated tickets. Discarding it overstates overall accuracy.

C:

The minimum is useful for a worst-group measure. It is not the requested accuracy across all evaluated tickets.

Question 60

A worker asked to optimize a function reports that it looks cleaner, but runtime performance is unchanged. What was missing from the assignment?

  1. A rule requiring fewer lines of code as the sole definition of optimization.
  2. A measurable success criterion tied to the required performance change.
  3. An instruction to maximize CPU utilization regardless of throughput or latency.
  4. A requirement to describe the patch in more detail without measuring execution.

Explanation

Correct Answer

B:

The task needed an observable target, such as benchmark improvement under fixed conditions. Aesthetic changes do not demonstrate that target.

Incorrect Answers

A:

A shorter implementation can be slower or unchanged in runtime. The assignment needs a performance measure tied to the user's goal.

C:

High utilization is not itself the requested performance improvement. Extra computation can increase resource use while leaving the function no faster.

D:

A more detailed narrative still provides no criterion for determining whether the optimization improved runtime behavior.

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