Anthropic CCA-F (Claude Certified Architect - Foundations (CCAR-F)) Practice Test 10

Question 90

A legal-research tool is advertised as providing definitive legal judgments, but it actually retrieves web snippets. What should its description say?

  1. Rename the tool to final_verdict without changing implementation
  2. State that retrieval does not certify legal conclusions
  3. Keep the authoritative claim because it increases tool usage
  4. Hide source locations to make the output appear cleaner

Explanation

Correct Answer

B:

Tool descriptions are the contract for model planning. A retrieval tool should represent evidence access accurately and preserve uncertainty.

Incorrect Answers

A:

Naming does not confer expert validation. The mismatch between capability and description would persist.

C:

Overclaiming capabilities can cause unsupported legal assertions. The agent may confuse retrieved text with verified advice.

D:

Removing citations makes claims harder to verify. Legal research especially needs provenance.

Question 91

A developer creates an export_records tool but omits information about the data it can expose. Which enhancement is most relevant?

  1. Remove the option to filter exported records
  2. Remove export event logging after the file is created
  3. Ask Claude to infer all privacy rules from prior casual chats
  4. Document export scope, sensitivity and side effects
  5. Publish raw export credentials in the tool description

Explanation

Correct Answer

D:

The agent must know when export is appropriate and what authorization is needed. Clear boundaries also support safer policy checks.

Incorrect Answers

A:

Broader exports increase unnecessary disclosure. Narrow filters help satisfy least-privilege data access.

B:

Suppressing logs would hide who exported sensitive information. Auditability and scope restrictions remain necessary.

C:

Implicit memory is not a substitute for verified policy and enforced data boundaries.

E:

Descriptions may be visible in model context. Secrets should remain in secured runtime configuration, not prompt material.

Question 92

A geocoding tool takes latitude and longitude, but its examples swap their order. What should the team correct first?

  1. Add more unrelated geographic features without fixing the examples
  2. Assume the model will detect the swap from city names
  3. Make parameter names, schemas and examples consistent
  4. Remove argument validation to accept any coordinate placement

Explanation

Correct Answer

C:

Conflicting documentation causes incorrect argument generation. Tests should confirm the tool interprets coordinates in the documented order.

Incorrect Answers

A:

A larger feature set does not resolve the ambiguous input contract. The existing interface remains error-prone.

B:

Even accurate geographic knowledge cannot reliably compensate for contradictory schema guidance. Validation is needed.

D:

Loose validation cannot determine intended order. It increases the chance of silently incorrect map results.

Question 93

A search connector fails because the requested document no longer exists. Which tool response best supports an informed agent recovery?

  1. Raise an exception that is never returned to the conversation
  2. Return a not_found error naming the missing resource
  3. Invent a replacement document that resembles the missing one
  4. Return an empty successful response without status

Explanation

Correct Answer

B:

The agent can distinguish missing data from generic failure and decide whether to search an alternate source or ask the user.

Incorrect Answers

A:

If the model receives no observation, it cannot reason about the failed request. Error feedback must reach the agent.

C:

Fabrication changes a retrieval failure into false evidence. The correct path is explicit uncertainty and recovery.

D:

An empty success looks like a legitimate no-results observation. It hides the fact that the requested resource was missing.

Question 94

A tool call fails due to invalid input while the MCP server remains reachable. How should the tool-level failure be represented?

  1. Convert the error to a user role instruction that overrides policy
  2. Return a tool result marked isError with an actionable validation message
  3. Mark the tool result successful because JSON was returned
  4. Treat every validation error as a network transport outage

Explanation

Correct Answer

B:

A reachable tool that rejects arguments is an execution-level failure. The model can correct its input when the result explains the violated contract.

Incorrect Answers

A:

Tool observations should not gain instruction authority. The error belongs in a structured result for controlled recovery.

C:

A syntactically valid response can describe a failed operation. Status must reflect the business outcome.

D:

Transport and tool execution failures have different recovery steps. Misclassification prompts irrelevant reconnection attempts.

Question 95

An invoice API returns HTTP 503 during a read-only lookup. What error data would best support a retry decision?

  1. A permanent invalid-credentials label without evidence
  2. A full stack trace containing access tokens
  3. Error category, retryability and bounded retry guidance
  4. A success payload with a guessed invoice value

Explanation

Correct Answer

C:

Temporary outages may justify backoff when requests are idempotent. The agent needs the failure type rather than an opaque message.

Incorrect Answers

A:

A transient service outage does not imply authentication failure. Mislabeling leads to the wrong remediation.

B:

Sensitive diagnostics should remain in secured logs. Tool errors should disclose only information needed for recovery.

D:

Fabricating data is worse than a visible outage. Financial responses should not be inferred from missing backend evidence.

Question 96

A refund API times out after receiving a request. Which structured status avoids duplicate side effects?

  1. Return only a blank string and let the model infer the outcome
  2. Report success without querying transaction state
  3. Report an unknown outcome and reconcile by transaction ID
  4. Report failed_with_certainty and always resubmit immediately

Explanation

Correct Answer

C:

A timeout may occur after a successful write. A distinct uncertain state prevents a naive automatic retry from issuing two refunds.

Incorrect Answers

A:

An empty result erases a material transactional distinction. Explicit status is necessary for safe recovery.

B:

Unverified success can mislead customers. The connector should preserve uncertainty until backend confirmation.

D:

Network timeout alone cannot establish that the server rolled back. Retries must inspect transaction state or use idempotency controls.

Question 97

A parser tool receives an unsupported file format. What should its error explain to the caller?

  1. Only the phrase something went wrong with no cause
  2. Name the unsupported format and available conversion paths
  3. A fabricated extracted record to keep the workflow moving
  4. The internal filesystem secrets of the parsing service

Explanation

Correct Answer

B:

Actionable input feedback lets the agent obtain a supported file or route to a converter. Generic failure text wastes retries.

Incorrect Answers

A:

Without a category or correction hint, the agent cannot choose an effective next step.

C:

Made-up extraction values contaminate downstream validation. Unsupported input should produce an honest failure.

D:

Secret paths and credentials do not help solve a format mismatch. They create unnecessary exposure.

Question 98

A model repeatedly calls a pricing tool with a missing required product ID. What error contract should the tool implement?

  1. Require the model to examine application logs unavailable to it
  2. Return zero as the price whenever an ID is missing
  3. Flag the invalid product_id field and its expected format
  4. Automatically substitute the first product in the catalog

Explanation

Correct Answer

C:

A precise message supports a corrected call or clarification. It is more useful than silent rejection and prevents unbounded guessing.

Incorrect Answers

A:

Private logs may help operators but do not give the agent an actionable result. Error messages need a safe public subset.

B:

A zero value looks like genuine commercial data. It can corrupt a downstream order or quote.

D:

Choosing a different item changes the request's meaning. The tool should not silently guess a business identifier.

Question 99

A tool returns isError=true but its accompanying content says record updated successfully. What is the primary defect?

  1. The response contains too many punctuation marks
  2. Conflicting status and payload make the outcome ambiguous
  3. The connector should hide the response entirely
  4. The model should ignore isError and trust optimistic prose

Explanation

Correct Answer

B:

The error flag and natural-language content must agree on whether the operation succeeded. Inconsistent contracts can cause incorrect retries or user claims.

Incorrect Answers

A:

Formatting details are secondary. A semantic conflict between result status and content changes the meaning of execution.

C:

Suppressing an inconsistent result prevents recovery. The interface should be corrected and return a truthful outcome.

D:

Status signals exist to communicate execution outcome. Ignoring them can turn failures into false success.

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.

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