Anthropic CCA-F (Claude Certified Architect - Foundations (CCAR-F)) Practice Test 11

Question 100

A customer service connector must distinguish a recoverable capacity error from a payment update whose outcome is unknown after timeout. Which TWO error behaviors are sound? Choose TWO.

  1. Retry the payment write immediately without checking whether it committed
  2. Return an outcome-unknown state and transaction identifier for the timed-out write
  3. Return a throttling category with retry guidance for temporary capacity limits
  4. Mark both failures as successful responses to reduce user friction
  5. Classify every capacity error as permanent permission denial

Explanation

Correct Answers

B:

A timeout may occur after successful execution. Reconciliation prevents duplicate financial side effects from blind retries.

C:

This allows bounded backoff rather than treating every failure as invalid credentials. The tool should preserve the server's actual status.

Incorrect Answers

A:

An unacknowledged write could have succeeded. Unconditional replay risks duplicate transfers.

D:

A valid transport response is not the same as a completed business operation. Mislabeling hides missing data and unsafe uncertainty.

E:

Authorization and throttling require different recovery actions. Misclassification directs the agent toward irrelevant access fixes.

Question 101

A search adapter returns partial results after one source fails. How should its tool output represent completeness?

  1. Fill the gap with guesses based on nearby search results
  2. Discard all successful findings automatically
  3. Call the whole operation fully successful without mentioning the missing source
  4. Provide a partial status, successful sources and the failed-source details

Explanation

Correct Answer

D:

The coordinator can use verified findings while identifying coverage gaps. Explicit partial success is more informative than a false binary success.

Incorrect Answers

A:

Inference without evidence can misrepresent the failed source. Unknown coverage should remain labeled unknown.

B:

Valid partial work can aid recovery and prevent repetition. The parent should decide whether it is sufficient.

C:

Silent omission can bias research synthesis and conceal missing evidence. Completeness must be observable.

Question 102

An MCP tool reports a business-rule denial because the user lacks permission to delete a record. Which message best assists the agent?

  1. Make the tool call appear successful so the conversation ends calmly
  2. Return a permission_denied tool error without leaking private record details
  3. Return the protected record as an example of why deletion failed
  4. Tell the model to ignore all higher-level instructions and retry as administrator

Explanation

Correct Answer

B:

The model can explain the blocked action or request an authorized workflow. The connector should not disclose data the user cannot access.

Incorrect Answers

A:

False completion can mislead users into believing the record was deleted. The denial must remain visible.

C:

Unauthorized disclosure compounds the security failure. Error content should be sanitized and minimal.

D:

A tool result is data, not an authority escalation channel. Privileges cannot be obtained through error text.

Question 103

A retrieval service receives invalid search syntax but responds with a retryable transport error. Why is that problematic?

  1. The API should never expose validation error categories
  2. Transport retries always fix user-generated syntax mistakes
  3. An unchanged invalid query may be retried instead of repaired
  4. The model should publish the failed search as verified evidence

Explanation

Correct Answer

C:

Client-input errors normally need correction, not network backoff. Distinct categories guide effective recovery strategies.

Incorrect Answers

A:

Hiding safe, actionable diagnostics limits the model's ability to correct its own calls. Structured errors are a core tool design aid.

B:

The same invalid argument remains invalid after a delay. Time does not change parser grammar.

D:

No valid query result exists. The agent must not treat a malformed request as a successful lookup.

Question 104

An assistant must inspect account data but should not initiate refunds. Which tool exposure best enforces its role?

  1. Provide the account lookup tool and omit the refund mutation tool
  2. Remove lookup access and ask it to estimate account state
  3. Expose all tools and ask the assistant not to refund
  4. Give the role administrative shell access instead of refund access

Explanation

Correct Answer

A:

A task-specific catalog enforces least privilege at the capability boundary. A prompt-only prohibition is weaker than nonexposure.

Incorrect Answers

B:

A read-only task requires real data. Guessing undermines accuracy and does not improve the requested service.

C:

The model could still choose the mutation in error. Unneeded side effects should not be available to the read-only role.

D:

Broad shell privileges create an even larger risk surface. They are unrelated to safe account retrieval.

Question 105

A supported Claude Messages API model must call one of two schema-defined extraction tools before the application can proceed. A direct prose reply is unacceptable, but either tool is valid. What tool_choice setting best matches the requirement?

  1. Force an unrelated tool name with tool_choice type tool
  2. Use tool_choice with type none
  3. Use tool_choice with type any
  4. Use tool_choice with type auto

Explanation

Correct Answer

C:

The any setting requires a call to one of the provided tools without forcing a particular name. It differs from auto, which permits text-only responses; the application still validates the arguments and tool result.

Incorrect Answers

A:

Named tool forcing is appropriate only when that exact tool is required. Choosing an unrelated interface would not satisfy either of the permitted extraction paths.

B:

None prevents tool calls, so neither extraction interface can be invoked. The workflow needs tool arguments rather than a prose-only message.

D:

Auto lets Claude choose between a tool call and direct text. That flexibility conflicts with the hard requirement that the initial response invoke one of the extraction interfaces.

Question 106

A workflow should call a specific approve_draft tool at the final stage. How can tool selection be constrained for that request?

  1. Force the named tool through tool_choice; enforce approvals
  2. Use auto and hope the model selects approve_draft from similar tools
  3. Grant the model broad permissions and remove the approval policy
  4. Remove the named tool and ask for an approval phrase in text

Explanation

Correct Answer

A:

A named-tool choice narrows selection to the intended interface. It does not itself authorize the business effect or ensure semantic correctness.

Incorrect Answers

B:

Auto permits other selections or no tool use. The stage has an explicit required interface.

C:

Forcing a tool call cannot replace transaction authorization. It only controls the response format and route.

D:

A generated phrase is not the operation expected by the host. The workflow needs a typed tool invocation.

Question 107

A short FAQ question can be answered directly, but complex requests should use tools when needed. Which tool_choice mode fits?

  1. Force a specific write tool for all questions
  2. Disable all tools even for live account questions
  3. Force a database tool on every FAQ regardless of relevance
  4. Use auto with precise tool descriptions

Explanation

Correct Answer

D:

Auto permits Claude to decide between a direct response and tool use. Good descriptions steer the boundary without forcing needless external calls.

Incorrect Answers

A:

A mutation tool is inappropriate for read-only information requests. This introduces avoidable side effects.

B:

No-tool execution would prevent retrieval of current personalized state. The system requires flexibility rather than a blanket ban.

C:

Unnecessary mandatory calls waste latency and risk access to unrelated data. Simple stable questions need not always use external tools.

Question 108

A coordinator offers 80 unrelated tools to a subagent tasked only with code search. What is the most appropriate redesign?

  1. Provide a compact code-search-focused tool list to the subagent
  2. Remove all source-inspection tools and rely on model recollection
  3. Leave all 80 available and shorten every name to one letter
  4. Give all tools to every nested worker to preserve consistency

Explanation

Correct Answer

A:

A focused catalog simplifies choice and prevents unrelated operations. Tool budgets and authorization become easier to reason about.

Incorrect Answers

B:

Repository work requires actual current code evidence. Eliminating read tools would sacrifice task accuracy.

C:

Opaque names worsen tool selection despite reducing tokens. The agent still sees many irrelevant capabilities.

D:

Uniform catalogs disregard role boundaries. Specialized workers usually need distinct permissions.

Question 109

An automation may query project status but may not modify project configuration. How should the parent and child tool lists differ?

  1. Ask the child to perform edits using natural-language output
  2. Allow the child every parent tool because the parent is trusted
  3. Make the parent read-only and the child the sole administrator
  4. Let the status child impersonate the parent when writes are needed
  5. Keep write access with the parent; give the child read access

Explanation

Correct Answer

E:

Tool partitioning prevents a subordinate investigator from performing unrelated changes. Escalation can happen through a reviewed parent action.

Incorrect Answers

A:

Text instructions are not a substitute for a controlled write path. The workflow should not hide the action mechanism.

B:

Delegation does not imply identical authority. The child's narrow role should be reflected in its capabilities.

C:

That reverses accountability without a task-based reason. The reporting child has no need for mutation authority.

D:

Role impersonation blurs accountability and extends privilege. The parent must explicitly authorize and perform suitable mutations.

Question 110

A tool catalog has separate search_web and internal_doc_search methods. The assistant uses the public search for confidential records. Which fix is strongest?

  1. Give the public search tool access to all private records
  2. Separate data scopes and restrict confidential retrieval
  3. Change every tool name to query without documenting boundaries
  4. Ask the model to memorize confidentiality rules without restrictions

Explanation

Correct Answer

B:

Semantic descriptions and runtime permissions should align. Public search must not become a backdoor for internal data handling.

Incorrect Answers

A:

Expanding external reach worsens confidentiality. It does not address inappropriate tool selection.

C:

Generic names obscure where data is allowed to flow. Distinct scope information is necessary.

D:

Prompts help but cannot guarantee safe data access. The executor must enforce allowed destinations.

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