Anthropic CCAR-P (Claude Certified Architect - Professional) Practice Test 2

Question 10

An engineering assistant receives thousands of lines of obsolete build logs in every request. Recent error evidence and relevant repository instructions are frequently omitted to fit the context limit. Which change should be tried first?

  1. Keep an equal token allowance from every historical build, truncating each proportionally
  2. Summarize only the old build logs and continue omitting current repository instructions
  3. Prioritize current diagnostic evidence and relevant instructions; retrieve older logs when needed
  4. Retain error-frequency counts from all builds and discard individual diagnostic messages

Explanation

Correct Answer

C:

The immediate failure is poor selection of information, not an inherent need to send every available record. Keeping high-value current evidence makes the limited input more useful for diagnosis.

Incorrect Answers

A:

Equal allocation treats stale and current evidence as equally valuable. It can still truncate the recent messages and relevant instructions that are needed for this failure.

B:

Compression can save space, but this selection still excludes information directly relevant to the current environment. Summarization is not a substitute for choosing the right evidence.

D:

Counts can identify recurring symptoms but discard stack traces, locations, and context. The assistant needs specific diagnostic evidence, not only an aggregate frequency report.

Question 11

A team expects cache reuse across requests, but a changing request identifier appears before otherwise identical long policy text. The requests are eligible for caching and arrive within the configured lifetime. Which change most directly supports reuse?

  1. Move the cache breakpoint later while retaining the identifier at the beginning
  2. Extend the cache lifetime while retaining the changing identifier before the policy
  3. Move the identifier after the stable prefix and its cache breakpoint
  4. Keep the input unchanged and retry requests more frequently within the cache lifetime

Explanation

Correct Answer

C:

Cache matching depends on the reusable prefix remaining consistent. Moving per-request content after that boundary preserves stable preceding material; usage metrics should then confirm actual cache reads.

Incorrect Answers

A:

A later breakpoint still includes the preceding changing identifier in the prefix being matched. The long policy text remains behind a difference that prevents that prefix from being reused.

B:

A longer lifetime helps a matching entry remain available, but does not make different prefixes identical. These requests already arrive within the configured lifetime.

D:

More frequent requests cannot remove the prefix difference. Retrying without restructuring the input repeats the same cause of the missed reuse.

Question 12

Eight teams share the same approved escalation rules but maintain separate prompt copies. Updates are applied inconsistently. Which design best improves consistency while preserving each team's task instructions?

  1. Distribute policy changes by email and let each team update its own prompt copy
  2. Compose team-specific task prompts from a shared, versioned, and reviewed policy module
  3. Load the latest policy dynamically without recording which revision was supplied
  4. Keep team-specific copies and compare their wording only after an escalation incident

Explanation

Correct Answer

B:

A common versioned component provides a reviewable update point, while separate task modules retain local requirements. Tests should cover the assembled prompts so a shared change does not silently break a team workflow.

Incorrect Answers

A:

Notification does not enforce adoption or record which prompt revision is active. Teams can still miss an update or interpret it differently.

C:

Dynamic retrieval may provide current policy, but unrecorded revisions make behavior difficult to audit and reproduce. Controlled composition needs traceable versions as well as shared content.

D:

Incident-triggered comparison detects drift after it has affected behavior. It does not provide a reliable shared update mechanism before deployment.

Question 13

A custom application sends Claude a client-side tool definition for querying shipment status. Claude returns a tool-use request with arguments. What must the application do next to obtain an evidence-based final answer?

  1. Execute the lookup but omit its result from the next request to reduce tokens
  2. Return the proposed lookup arguments as a successful tool-result payload
  3. Validate, execute, and return the result with its tool-use identifier
  4. Wait for a server-tool completion event before running the custom client operation

Explanation

Correct Answer

C:

For a client-side tool, application code performs the actual operation. Returning the result with the matching tool-use identifier lets the subsequent model response use the retrieved evidence.

Incorrect Answers

A:

Execution alone does not put the carrier response into the model context. The following answer cannot use evidence that the application has not returned.

B:

The arguments describe the proposed operation, not the carrier's response. Using them as evidence would confuse a requested action with a completed lookup.

D:

The declared operation is a client-side tool, so the application supplies its implementation. Waiting for server-side execution does not complete this custom lookup.

Question 14

A validly authenticated employee asks a record-search tool for another tenant's account. The generated tool arguments contain that tenant's identifier. Which control must decide whether the lookup can proceed?

  1. Check the caller’s permitted tenant scope at the service boundary
  2. Validate that the tenant identifier matches the tool’s JSON schema
  3. Accept the selected tenant if it exists in the organization’s account directory
  4. Compare the tenant identifier against a natural-language instruction in the prompt

Explanation

Correct Answer

A:

The service can compare the authenticated principal with the requested resource and deny an out-of-scope lookup. Authorization must not depend on whether the model generated convincing arguments.

Incorrect Answers

B:

Schema validation can establish that an identifier has the expected type or format. It does not prove that the caller may access the named tenant.

C:

Resource existence establishes that the account is real, not that this employee may access it. The lookup needs an entitlement check tied to the caller.

D:

Prompt guidance can steer model behavior but does not enforce an access decision. The data service must reject unauthorized requests regardless of their source.

Question 15

An HTTP MCP server receives a bearer token intended for a different protected service. The token is well formed and not expired. What should the MCP server do?

  1. Accept the token when its signature, issuer, and expiry checks pass
  2. Accept the token when its scopes name the requested operation
  3. Reject the token because its audience is a different resource
  4. Forward the token to the upstream resource and reuse that resource’s acceptance decision

Explanation

Correct Answer

C:

The server must validate that the access token is intended for its resource, alongside other required checks. Correct formatting and lifetime do not fix an audience mismatch.

Incorrect Answers

A:

A trusted issuer and valid signature establish aspects of token authenticity. They do not change the resource for which the token was issued.

B:

A scope name is interpreted within its authorization context. A matching-looking scope on a token for another audience does not authorize this server.

D:

Another resource may legitimately accept a token intended for it. That decision does not transfer authorization to this MCP server, which must validate its own audience binding.

Question 16

An asset assistant must find the exact part number ZX-417-Q even when semantically similar manuals discuss different components. Semantic search alone often ranks a different part first. Which retrieval adjustment is best justified?

  1. Use a semantic similarity cutoff and select the first passage above it
  2. Combine exact-identifier lexical matching with semantic retrieval for descriptive fault queries
  3. Normalize away identifier punctuation and all digits before indexing the manuals
  4. Expand the query into descriptive synonyms and rank solely by embedding similarity

Explanation

Correct Answer

B:

Exact-string evidence and semantic similarity solve different retrieval needs. Combining them can preserve precise identifier matching while still supporting natural-language descriptions of a fault.

Incorrect Answers

A:

A similarity threshold can remove weak matches, but a different component may still be semantically very close. It does not enforce the exact part-number distinction.

C:

Normalization must preserve meaningful identity. Removing the digits would merge distinctions between related components and undermine this exact-match requirement.

D:

Query expansion can improve recall for descriptive searches but can also favor related components. The missing signal is precise identifier matching, not more descriptive similarity.

Question 17

An indexed maintenance table is split so that numeric limits are separated from their column headers and units. Claude retrieves the numbers but assigns them to the wrong operating conditions. What ingestion change should be evaluated first?

  1. Use smaller numeric-only chunks to make each retrieved value more specific
  2. Preserve row, header, and unit context in each retrievable table segment
  3. Keep the existing chunks and increase vector-index search precision
  4. Keep isolated cells and rerank them using their numeric values as the main signal

Explanation

Correct Answer

B:

A retrieved unit should contain enough information to interpret its values correctly. Keeping table structure or adding concise contextual metadata can prevent a numeric cell from being treated as an isolated fact.

Incorrect Answers

A:

A smaller cell fragment may be more focused but loses the relationships that define its meaning. Isolating a value cannot restore its missing header or unit.

C:

More precise vector search can improve ranking of existing representations. It cannot reliably recover structure that ingestion already discarded.

D:

A number alone does not identify the relevant operating condition. Reranking contextless cells retains the same ambiguity even if their order changes.

Question 18

Total request logs show occasional 20-second responses. Separate model-service metrics suggest typical calls take under 3 seconds, but there are no correlated stage timings. Which addition best determines where slow requests spend their time?

  1. Trace correlated durations for queueing, retrieval, tools, models, and retries
  2. Increase metric sampling for total latency without recording individual pipeline stages
  3. Collect separate service averages without correlating them to a request
  4. Record complete final answers and the time when each answer was returned

Explanation

Correct Answer

A:

A shared request identifier and stage spans show where time was spent on the same transaction. This distinguishes external dependency delays from model generation or application queueing.

Incorrect Answers

B:

More samples improve the estimate of the total distribution, but still do not show where a particular request waited. The missing dimension is stage-level timing.

C:

Independent averages can hide whether delays occur together on the same slow transaction. Correlated spans are needed to reconstruct that transaction’s critical path.

D:

Final-answer records provide a completion event but no breakdown of prior processing. They cannot separate a queue delay from a slow tool or retry.

Question 19

A carrier supports operation-key idempotency and status lookup within a documented retention window. A shipping-label tool times out after the carrier may have accepted the request; the operation key is still valid. Which recovery design avoids duplicate charges while completing the order?

  1. Retry the timed-out operation with a fresh operation key and exponential backoff
  2. Assume the timeout means rollback and issue the same payload again
  3. Query operation status and retry safely with the original retained idempotency key
  4. Record success locally after the timeout and continue the shipping workflow

Explanation

Correct Answer

C:

The provider’s status endpoint can resolve an uncertain result, and its idempotency guarantee lets a repeated request refer to the same business operation. The retry must preserve the original key and comply with the provider’s retention and payload rules.

Incorrect Answers

A:

A fresh identifier makes a retry look like a new business operation. It can defeat duplicate detection and produce the second charge the design must prevent.

B:

A transport timeout does not reveal whether the carrier committed the operation. Treating it as rollback can create a second billed label when the first actually succeeded.

D:

Recording success without evidence can leave the order pointing to a label that does not exist. Ambiguous completion requires reconciliation rather than inventing either success or failure.

Question 20

A platform exposes 600 internal tools, but each support task uses only a few. Loading every detailed schema consumes most input tokens and confuses selection. Which design should be evaluated?

  1. Cache the complete tool catalog but still expose every schema for each decision
  2. Raise the context limit while retaining all detailed schemas in every request
  3. Use task-directed tool discovery and load the selected detailed interfaces only when needed
  4. Rotate through fixed catalog batches without matching them to the current task

Explanation

Correct Answer

C:

A discovery step narrows the active interface set before execution. It reduces irrelevant input while preserving access to capabilities when the task actually requires them.

Incorrect Answers

A:

Caching can reduce repeated input processing costs, but does not narrow the set of interfaces the model must distinguish. The tool-selection confusion remains.

B:

A larger context may fit the catalog without addressing its relevance or the ambiguity between similar tools. Capacity alone does not solve capability bloat.

D:

Fixed batches can exclude the needed interface or require several additional searches. Task-directed discovery is better matched to selecting a small relevant subset.

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