Anthropic CCAR-P (Claude Certified Architect - Professional) Practice Test 12

Question 110

A prompt mixes developer directions and a quoted customer email without clear boundaries. Which revision best improves interpretation while retaining separate security controls?

  1. Mark the entire customer email as a system directive
  2. Remove all labels and rely on paragraph order
  3. Use delimiters as the only security enforcement
  4. Clearly label the task instructions and quoted source data

Explanation

Correct Answer

D:

Explicit boundaries help the model distinguish what to do from what to analyze. They improve clarity but do not replace authorization or injection defenses.

Incorrect Answers

A:

This promotes untrusted content into apparent authority. It makes the original distinction harder rather than clearer for the model.

B:

Position alone can be ambiguous when source material contains commands. Labels provide a clearer statement of each section's purpose.

C:

Formatting can guide interpretation but cannot guarantee safe behavior. Consequential actions still need independent application controls.

Question 111

A shared prompt says to summarize in two sentences, while a task module requires five separately labeled findings. Outputs vary. What should the prompt owner do?

  1. Repeat both instructions more often
  2. Resolve the conflict into one explicit output requirement
  3. Concatenate both conflicting rules and leave their order unchanged
  4. Choose whichever output appears more polished

Explanation

Correct Answer

B:

Incompatible requirements force the model to improvise a priority. A coherent contract removes the ambiguity instead of depending on which instruction receives attention.

Incorrect Answers

A:

Repetition does not make incompatible requirements simultaneously satisfiable. It can amplify confusion while consuming additional context.

C:

Including both rules without resolving their incompatibility leaves the model no consistent target. The prompt needs an explicit agreed output requirement.

D:

Post hoc preference hides the unstable specification. The owner should define the intended behavior before evaluating compliance.

Question 112

Few-shot examples accidentally label refund requests as delivery questions. The classifier now repeats that mistake. What should be corrected first?

  1. The retrieval ranking used for unrelated background passages
  2. The mislabeled demonstrations and their review process
  3. The formatting of the demonstrations without checking their labels
  4. The number of demonstrations, increased using the same labeling process

Explanation

Correct Answer

B:

Examples provide concrete guidance about category boundaries. Incorrect labels can teach the wrong mapping even when the general instruction is accurate.

Incorrect Answers

A:

Retrieval ranking can affect supplied evidence but does not correct demonstrations that explicitly teach the wrong labels. The known labeling defect should be repaired first.

C:

Presentation can affect readability but does not repair examples that teach the wrong classification. Their labeled outcomes need substantive review.

D:

Scaling a defective labeling process can reinforce the wrong behavior. The demonstrated classification rule must be corrected before adding more examples.

Question 113

Two unrelated case files are analyzed in one request. Each file is labeled, but the answer attributes a date from case B to case A. Which prompt contract most directly addresses this association error?

  1. Merge both timelines before assigning conclusions to the cases
  2. Require one overall conclusion supported by any supplied passage
  3. Sort all extracted dates chronologically without retaining case labels
  4. Require case-keyed results with evidence confined to the corresponding file

Explanation

Correct Answer

D:

Stable case identifiers connect each result to its permitted evidence. Validate those associations so a plausible date from another file cannot silently support the result.

Incorrect Answers

A:

Merging first discards the boundary that matters for this task. A coherent combined timeline can still assign an event to the wrong case.

B:

A shared evidence pool permits exactly the cross-case attribution observed. Each case needs conclusions supported by that case's material.

C:

Chronological order does not preserve ownership of an event. Removing case labels makes later attribution harder to verify.

Question 114

A user edits examples inside one request and expects the underlying model weights to be permanently updated for all future calls. Which explanation is accurate?

  1. The examples guide that context without themselves retraining weights
  2. The examples have no effect unless weights are changed
  3. Every accepted example changes the deployed model weights
  4. Future calls inherit the examples without storing or resending them

Explanation

Correct Answer

A:

Prompted examples influence the current request's behavior. They do not automatically become a persistent model update across unrelated future calls.

Incorrect Answers

B:

Models can adapt their responses to demonstrations within context. Persistent training is not required for examples to guide an individual task.

C:

That confuses inference-time context with a training operation. Supplying text in a request does not by itself perform model retraining.

D:

Persistence needs an explicit mechanism in the application or product. A standalone request does not guarantee that later requests receive its examples.

Question 115

An extractor must use null when a field is absent, but examples always supply a value. What prompt change best teaches the missing-field behavior?

  1. Add more fully populated examples
  2. Include absent-field cases and define null precisely
  3. Replace null with a plausible default value
  4. Ask for confident answers without uncertainty markers

Explanation

Correct Answer

B:

The model needs both the rule and demonstrations of when to apply it. Examples should show that unsupported values remain absent rather than guessed.

Incorrect Answers

A:

Complete records never demonstrate the required missing-field branch. They can reinforce the expectation that every output field needs a substantive value.

C:

A default can be useful only if the business contract permits it. Here the required behavior is explicitly to represent absence.

D:

Confidence-oriented wording can encourage completion of unsupported fields. It conflicts with the requirement to preserve missing information honestly.

Question 116

A research summary must let readers verify each material claim against supplied passages. Which prompt requirement best supports that goal?

  1. Cite the first passage for every sentence
  2. Associate claims with valid supplied evidence identifiers
  3. Allow invented source names when no identifier fits
  4. Put a list of all documents after the summary

Explanation

Correct Answer

B:

Claim-level references expose the relationship between conclusions and sources. The application should also check that identifiers exist and actually support the claim.

Incorrect Answers

A:

A valid identifier can still be irrelevant. Citation quality requires support for the particular statement, not merely an existing source label.

C:

Fabricated references make verification impossible. The model should instead indicate that the claim lacks support or omit it under the task rules.

D:

A bibliography identifies available material but not which passage supports each claim. It can obscure unsupported conclusions among many real sources.

Question 117

An analyst needs both reported facts and reasoned implications from a document. How should the prompt structure the output?

  1. Mix quotations and hypotheses without labels
  2. Remove all reasoning and return only document titles
  3. Present every implication as a direct statement from the document
  4. Separate source-supported facts from explicitly labeled inferences

Explanation

Correct Answer

D:

The reader can then distinguish what the document says from what the analyst concludes. This preserves useful reasoning without presenting it as quoted evidence.

Incorrect Answers

A:

Unlabeled mixing makes verification difficult. Readers may treat a speculative implication as an authoritative statement from the document.

B:

Titles do not provide the requested factual and analytical content. The requirement permits inference when its status is clearly distinguished.

C:

That misattributes the assistant's reasoning to the source. Even a reasonable inference is not the same as an explicitly reported fact.

Question 118

A prompt contains dozens of rare exceptions but omits the ordinary decision rule. The model handles common cases inconsistently. What revision is best?

  1. Add more rare exceptions before defining the rule
  2. Remove every exception regardless of consequence
  3. State the core rule clearly and organize relevant exceptions
  4. Use a longer role description instead of decision criteria

Explanation

Correct Answer

C:

A clear default gives the exceptions a meaningful reference point. Organizing them by applicability reduces confusion without discarding legitimate constraints.

Incorrect Answers

A:

More detail cannot compensate for a missing governing principle. It may further obscure the common behavior the system needs most often.

B:

Some exceptions may be essential to correctness. The remedy is coherent organization, not indiscriminate deletion of valid requirements.

D:

A persona can shape style but does not supply the operational rule. The model needs concrete criteria for making the required decision.

Question 119

A long-context answerer performs inconsistently when key evidence is buried among many irrelevant passages. What is the best next step?

  1. Assume any text inside the context is used equally well
  2. Test relevance filtering and evidence placement on controlled cases
  3. Increase the unrelated passages to reinforce completeness
  4. Lower the output limit until answers become shorter

Explanation

Correct Answer

B:

The issue may concern how information is presented rather than its mere presence. A controlled comparison can identify a useful ordering and selection strategy.

Incorrect Answers

A:

Being present does not guarantee equal influence on the response. The observed inconsistency is evidence that presentation deserves evaluation.

C:

Additional noise can make the relevant evidence harder to use. Completeness should be judged by task coverage rather than raw context volume.

D:

Concise output does not establish better evidence use. The intervention should target the input-selection and placement problem being observed.

Question 120

A long task is summarized to save context, but the summary omits a restriction against contacting external recipients. What should the compaction design preserve explicitly?

  1. Active constraints, decisions, pending work, and evidence references
  2. Every generated thought but no explicit task restrictions
  3. Only the chronological sequence of tool names
  4. Only the most recent user sentence

Explanation

Correct Answer

A:

Compaction should retain the information that governs future behavior. Losing a live restriction can change the task's authority even if the summary captures its topic.

Incorrect Answers

B:

Internal verbosity does not substitute for the operational state. The next step needs the actual requirements and boundaries that remain in force.

C:

A tool list can omit the purpose, constraints, and unresolved decisions. It is not a sufficient representation of the state needed to continue safely.

D:

Earlier instructions may still apply. Recency alone is not a reliable criterion for deciding which constraints survive compression.

Claude Certified Architect - Professional (CCAR-P) Practice Tests