Anthropic CCAR-P (Claude Certified Architect - Professional) Practice Test 13

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.

Question 121

A compacted conversation states that a vendor was approved, but the approval conditions were lost. What should the assistant do before relying on that approval?

  1. Treat the short summary as unconditional approval
  2. Discard all historical decisions permanently
  3. Retrieve the authoritative decision record and its conditions
  4. Ask a model to invent likely approval conditions

Explanation

Correct Answer

C:

A summary can omit qualifiers that materially change a decision. Recovering the original record prevents a compressed statement from becoming broader authority.

Incorrect Answers

A:

This expands the decision beyond what the available evidence establishes. Missing conditions should not be interpreted as absence of conditions.

B:

The problem is incomplete retrieval, not proof that every past decision is unusable. Recovering the relevant record preserves valid prior work.

D:

Plausible reconstruction is not evidence of what was agreed. The decision needs its actual recorded basis before it can govern action.

Question 122

A policy answer depends on a rule in one document and an exception in another. Each fits individually, but retrieval supplies only one. Which context strategy fits?

  1. Summarize both titles without their contents
  2. Alternate documents between repeated requests
  3. Assemble both relevant passages with their applicability metadata
  4. Choose whichever passage has the highest standalone similarity

Explanation

Correct Answer

C:

The decision requires combining the rule and exception. Supplying only one can yield a confident answer that ignores the other governing condition.

Incorrect Answers

A:

Titles do not establish the detailed rule or exception. The model needs the substantive passages and conditions that determine applicability.

B:

Alternation gives each call incomplete evidence and produces unstable answers. The required relationship should be considered within the same decision context.

D:

Similarity ranking alone may omit a necessary complementary source. The task requires coverage of both components, not a single best-looking passage.

Question 123

A reusable prefix is expensive to write into cache and cheaper to read. A workload almost never repeats a prefix. What must be checked before claiming savings?

  1. A cache hit proves the answer itself was reused
  2. A cost estimate that counts cache reads but excludes initial cache writes
  3. Every cache write automatically reduces total cost
  4. A longer prefix guarantees a higher hit rate
  5. Expected reuse must offset the cache-write overhead

Explanation

Correct Answer

E:

Caching has an economic benefit only when eligible reuse occurs often enough. A low-hit workload can fail to recover the initial write cost.

Incorrect Answers

A:

Input-prefix reuse does not imply response reuse. The model can still generate a different answer from the cached context and new suffix.

B:

A read-only estimate understates cost when reuse is low. The comparison must include the initial write overhead as well as subsequent reads.

C:

A write is not the same as a discounted read. The comparison must include the actual workload's reuse frequency and applicable prices.

D:

Prefix length affects potential processing saved, not whether later prefixes match. A long changing prefix may rarely be reused.

Question 124

A team enables prompt caching and expects identical answers to be returned without generation. Which distinction should be clarified?

  1. Answer caching needs no task-equivalence checks
  2. A cache hit guarantees deterministic wording
  3. Prompt caching reuses input processing, not a stored final answer
  4. Prompt caching permanently stores every generated answer

Explanation

Correct Answer

C:

The model still generates a response using the request context. A separate application response cache has different matching, freshness, and authorization requirements.

Incorrect Answers

A:

An application response cache must determine whether reuse is valid. Different users, evidence versions, or questions can require different answers.

B:

Reusing an input prefix does not make generation inherently deterministic. Other request content and sampling behavior can still affect the answer.

D:

That is not the input-prefix mechanism being configured. Treating it as a response store leads to incorrect cost and behavior expectations.

Question 125

Repeated requests use the same explicitly marked prefix and arrive within its configured cache lifetime. Cache creation and read token counts remain zero. The prefix is shorter than the selected model's documented minimum. What explains the result?

  1. The first cache read must occur before a cache write can be recorded
  2. The prefix is ineligible because it is below the applicable minimum length
  3. The generated answer must be longer than the input prefix to enable caching
  4. The cache lifetime must expire once before this prefix can be cached

Explanation

Correct Answer

B:

An explicit cache marker does not waive the minimum prefix length. Check the documented requirement for the selected model and deployment, then use usage counters to confirm caching.

Incorrect Answers

A:

A cache entry must be created before it can be reused. A prefix below the supported minimum does not become eligible by reversing that sequence.

C:

Prompt caching concerns eligible input prefixes. Generated answer length does not satisfy a minimum length requirement for cached input.

D:

Expiry removes reusable state rather than making a short prefix eligible. The requests are already timely; the stated problem is the prefix length.

Question 126

A hypothetical workload sends 10,000 input tokens per request. Usage records show 7,000 are cache reads and 3,000 require uncached processing. What fraction of input tokens received cache-read treatment?

  1. 30%
  2. 70%
  3. 100%, because every request contains some cached tokens
  4. A 70% reduction in the total API bill

Explanation

Correct Answer

B:

The token-level fraction is 7,000 divided by 10,000. This measures cached input-token share, not the fraction of requests or the exact reduction in total bill.

Incorrect Answers

A:

That is the uncached share of the input. It describes the complementary portion rather than the input tokens handled as cache reads.

C:

A request containing a cache hit can still include uncached input. Request-level hit presence does not mean every input token receives cached treatment.

D:

The token share alone does not determine overall monetary savings. Cache pricing, writes, uncached processing, and output charges also affect the bill.

Question 127

A rendered prompt combines policy, task instructions, and examples from separate repositories. What should a release record capture?

  1. Only the filename of the final template
  2. Only the most recently edited component's revision
  3. The exact versions of all assembled components
  4. Only the application server's operating-system version

Explanation

Correct Answer

C:

Component-level provenance makes the rendered behavior reconstructable. Recording only the outer template can miss changes in an imported policy or example set.

Incorrect Answers

A:

A filename can remain unchanged while its contents or dependencies change. It is insufficient to identify what was actually supplied.

B:

An assembled prompt can combine several independently versioned parts. Recording one changed component does not identify the complete configuration used for the request.

D:

Infrastructure version can aid debugging but does not identify prompt semantics. The relevant input components still need their own provenance.

Question 128

A team compares two prompt changes using one generated answer from each. Their model configuration produces variable outputs. What improvement is needed?

  1. Change the test questions between every trial
  2. Repeat controlled trials and compare outcome distributions
  3. Select whichever one answer looks more fluent
  4. Assume fixed input guarantees identical output

Explanation

Correct Answer

B:

A single sample can reflect random variation rather than a prompt effect. Repeated comparable trials reveal reliability and the spread of outcomes.

Incorrect Answers

A:

Changing inputs and prompts together makes differences harder to attribute. Controlled repetition needs comparable tasks across candidate configurations.

C:

Fluency in one sample is weak evidence of reliable improvement. It can favor a lucky output while hiding recurring failures.

D:

Identical input does not always imply deterministic generation. The evaluation design must reflect the behavior of the actual configuration.

Question 129

A configuration offers a tunable reasoning-effort control. Increasing it improves complex cases but raises latency. Which deployment policy is best justified?

  1. Use minimum effort because all prompts contain the same instructions
  2. Use the maximum effort for every request without measurement
  3. Evaluate effort levels by task complexity and service constraints
  4. Treat reasoning effort as an authorization control

Explanation

Correct Answer

C:

Additional reasoning should earn its cost through measured outcomes. Different tasks may need different settings to meet quality and response-time requirements.

Incorrect Answers

A:

Instruction similarity does not establish equal reasoning difficulty. Complex cases can require more processing to satisfy the same business contract.

B:

More computation can be unnecessary for simple tasks and may violate latency goals. The appropriate setting depends on observed task performance.

D:

Effort affects computation, not permission to access data or perform actions. Security boundaries must be enforced independently of the model's reasoning budget.

Question 130

A customer wants confidence in a financial answer without receiving private internal reasoning. Which output contract is most useful?

  1. Return the result, evidence, assumptions, and checkable calculations
  2. Remove all supporting calculations to keep the answer concise
  3. Return only an unexplained confidence percentage
  4. Require every hidden internal token to prove accuracy

Explanation

Correct Answer

A:

Verification can rely on observable support rather than a private reasoning transcript. Clear assumptions and reproducible calculations make the answer assessable.

Incorrect Answers

B:

Concision should not eliminate the material needed to check a financial conclusion. The contract can be compact while preserving the calculation basis.

C:

A percentage alone does not show whether inputs or calculations are valid. It provides less verifiable support than evidence and explicit assumptions.

D:

A reasoning transcript is not necessary or sufficient for correctness. The customer needs evidence that can be checked against the task and source data.

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