Anthropic CCDV-F (Claude Certified Developer - Foundations) Practice Test 38

Question 370

A retrieved PDF embeds instructions demanding a secret transfer unrelated to the user's question. Which design would meet this acceptance criterion?

  1. Label external text as untrusted input
  2. Attach source IDs to retrieved excerpts
  3. Use an external source refresh for changing facts
  4. Track source timestamps or versions

Explanation

Correct Answer

A:

Trust-boundary marking prevents source content from being promoted to application authority; the implementation must still recognize that labels should be backed by tool restrictions because prompts can be attacked.

Incorrect Answers

B:

Citation metadata lets downstream logic connect statements to original retrieved material; the missing capability is more specific: trust-boundary marking prevents source content from being promoted to application authority; without accurate source mapping, fabricated references can look convincing.

C:

Refetching authoritative information avoids treating outdated context as eternally current; by comparison, the intended solution relies on this: trust-boundary marking prevents source content from being promoted to application authority; the model should not invent a change simply because its history is old.

D:

Version metadata allows the agent to prefer current evidence and flag historical material; the needed capability is different—trust-boundary marking prevents source content from being promoted to application authority—and timestamp alone does not establish which source is authoritative.

Question 371

A release changes system instructions and reviewers need a diff and rollback target. What is the correct engineering decision?

  1. Protect API keys through deployment secrets
  2. Run contract tests against a versioned API schema
  3. Use code review on tool schema changes
  4. Commit prompt templates as versioned files

Explanation

Correct Answer

D:

Version control makes prompt changes auditable alongside application code and tests; its operational boundary is that storing production prompts only in an engineer's notebook prevents reproducibility.

Incorrect Answers

A:

Secret injection provides runtime access without publishing keys in code or artifacts; however, the critical mechanism here is that version control makes prompt changes auditable alongside application code and tests, and obscuring a key in a code comment is not adequate credential protection.

B:

Contract tests enforce agreed field names, types, and error shapes across producer and client; the requirement instead turns on this behavior: version control makes prompt changes auditable alongside application code and tests; casual manual testing may miss a breaking field change.

C:

Reviewing the schema and authorization path catches privilege expansion before release; this cannot substitute for the behavior that version control makes prompt changes auditable alongside application code and tests, because merely changing the tool description does not guarantee safe backend behavior.

Question 372

After a prompt change, agents attempt unexpected outbound requests more frequently. Which option best fits the described workflow?

  1. Use a policy-based refusal and safe redirect
  2. Monitor and investigate unusual tool activity
  3. Enforce least-privilege infrastructure permissions
  4. Canary a risky policy change

Explanation

Correct Answer

B:

Telemetry and alerts identify suspicious action patterns and allow prompt, tool, or policy remediation; notably, monitoring should complement, not replace, preventive execution controls.

Incorrect Answers

A:

A policy layer handles the request appropriately and communicates a safe boundary; the requirement instead turns on this behavior: telemetry and alerts identify suspicious action patterns and allow prompt, tool, or policy remediation; policy adherence should be evaluated; it is not guaranteed by a single warning prompt.

C:

Backend credentials and IAM policies restrict it to the minimum necessary operations; in contrast, telemetry and alerts identify suspicious action patterns and allow prompt, tool, or policy remediation, while an unrestricted credential remains dangerous regardless of the model's stated intent.

D:

A limited rollout measures refusal, error, and unauthorized action attempts before broad exposure; this addresses a separate concern, not the fact that telemetry and alerts identify suspicious action patterns and allow prompt, tool, or policy remediation; immediate all-user changes make regressions harder to contain.

Question 373

A system-prompt edit might change extraction accuracy despite no code changes. Which production safeguard is most relevant?

  1. Use regression evaluations for prompts
  2. Run an end-to-end test using a sandbox
  3. Use automated integration tests before release
  4. Perform security abuse tests

Explanation

Correct Answer

A:

A frozen dataset and grading rubric compare model outputs before and after the proposed prompt revision; the important constraint is that passing a linter does not measure semantic output quality.

Incorrect Answers

B:

An end-to-end test exercises the public request path, model call, dispatcher, and response formatter; the requirement instead turns on this behavior: a frozen dataset and grading rubric compare model outputs before and after the proposed prompt revision; isolated component success cannot prove the integration as a whole functions.

C:

Integration tests check how independently coded components actually interact under realistic conditions; this cannot substitute for the behavior that a frozen dataset and grading rubric compare model outputs before and after the proposed prompt revision, because unit tests of isolated helpers cannot reveal all cross-component contract failures.

D:

Red-team cases test injection resistance and authorization enforcement at trust boundaries; in contrast, a frozen dataset and grading rubric compare model outputs before and after the proposed prompt revision, while ordinary happy-path tests will not expose deliberate instruction attacks.

Question 374

A deployed workflow depends on stable tool-choice behavior while a model version is changed. Which of the following would work as specified?

  1. Refresh prompt evaluations after upgrade
  2. Recheck behavior after version migration
  3. Revalidate tool schemas and outcomes
  4. Update cost and context assumptions

Explanation

Correct Answer

C:

Integration tests inspect requested tool names, arguments, and final answer handling, although syntax compatibility does not guarantee identical agent decisions.

Incorrect Answers

A:

Re-running the fixed evaluation set detects behavior shifts requiring prompt adjustment; by comparison, the intended solution relies on this: integration tests inspect requested tool names, arguments, and final answer handling; copying old pass rates forward would misstate new performance.

B:

Regression tests evaluate prompts, tool arguments, latency, and output shape on the new version; the needed capability is different—integration tests inspect requested tool names, arguments, and final answer handling—and version changes can affect behavior even when the API request remains syntactically valid.

D:

Recalculating budgets avoids relying on stale token or feature assumptions; however, the critical mechanism here is that integration tests inspect requested tool names, arguments, and final answer handling, and product names alone do not prove pricing or capacity parity.

Question 375

An application fails but the team cannot tell whether its wrapper or the Claude service is responsible. What is the correct engineering decision?

  1. Trace the external tool service
  2. Reproduce with a raw valid API request
  3. Inspect the serialized outgoing payload
  4. Treat 400-class validation responses as request defects

Explanation

Correct Answer

B:

An isolated direct call distinguishes transport and wrapper behavior from a more complex integration path; in production, a successful API response does not automatically validate downstream application logic.

Incorrect Answers

A:

A distributed trace follows the tool dispatcher to the failing downstream dependency; the missing capability is more specific: an isolated direct call distinguishes transport and wrapper behavior from a more complex integration path; rewriting system instructions will not repair an unavailable invoice endpoint.

C:

Logging a redacted request reveals malformed fields introduced by serialization or adapters; the needed capability is different—an isolated direct call distinguishes transport and wrapper behavior from a more complex integration path—and reading the model's final natural-language answer cannot reveal a request that never reached generation.

D:

The developer checks the request schema, supported fields, and types before attempting another call, whereas the requirement calls for another behavior: an isolated direct call distinguishes transport and wrapper behavior from a more complex integration path; retrying an identical malformed request wastes time and capacity.

Question 376

A tool must learn a nonobvious output convention for mapping customer statuses. What should the team do at this decision point?

  1. Evaluate examples for data leakage
  2. Use multiple diverse examples for edge cases
  3. Use one representative example to clarify format
  4. Count the complete request against context limits

Explanation

Correct Answer

C:

A single illustrative input-output pair can demonstrate the intended pattern, although one example may overfit to a narrow case and should be tested on alternatives.

Incorrect Answers

A:

Approved synthetic or sanitized cases teach the format without disclosing unrelated private content; by comparison, the intended solution relies on this: a single illustrative input-output pair can demonstrate the intended pattern; good few-shot performance does not justify embedding sensitive identifiers.

B:

Contrasting examples illustrate boundaries and exceptions across several realistic inputs; the needed capability is different—a single illustrative input-output pair can demonstrate the intended pattern—and few-shot prompts add tokens and can create bias if examples are unrepresentative.

D:

The context budget encompasses the relevant input contents and model-specific output allowances; however, the critical mechanism here is that a single illustrative input-output pair can demonstrate the intended pattern, and counting only the latest user text understates true token pressure.

Question 377

Two tool requests try to update the same account balance in one interaction. Which solution matches the actual constraint?

  1. Require explicit user confirmation before expensive writes
  2. Serialize conflicting writes to shared state
  3. Execute dependent calls in the correct sequence
  4. Dispatch every read-only call immediately without a concurrency limit

Explanation

Correct Answer

B:

Coordination or transactions prevent lost updates and inconsistent final state; its operational boundary is that independent model-generated reasoning does not provide concurrency control.

Incorrect Answers

A:

An enforced approval step separates informative recommendations from transactional execution; this cannot substitute for the behavior that coordination or transactions prevent lost updates and inconsistent final state, because a prompt that says ask first cannot replace backend commit gating.

C:

A follow-up action depends on the actual result of the first tool operation; this addresses a separate concern, not the fact that coordination or transactions prevent lost updates and inconsistent final state; running both in parallel can reference an object that does not yet exist.

D:

Unbounded fan-out starts all requested reads without regard to supplier throughput; here, the controlling capability is that coordination or transactions prevent lost updates and inconsistent final state, but independent requests can still trigger quotas or upstream overload when no concurrency ceiling is enforced.

Question 378

An invoice tool receives an array of line items, each requiring a product ID and bounded quantity. What is the most direct implementation choice?

  1. Mark truly required arguments in the schema
  2. Provide precise tool descriptions and argument types
  3. Constrain nested object shapes
  4. Serialize conflicting writes to shared state

Explanation

Correct Answer

C:

Structured nested validation rejects incomplete or malformed records consistently—with this limitation: accepting arbitrary object shapes can produce unpredictable writes.

Incorrect Answers

A:

The schema identifies mandatory fields so calls missing them can be rejected before dispatch, whereas the requirement calls for another behavior: structured nested validation rejects incomplete or malformed records consistently; optional property defaults should not hide missing business-critical identifiers.

B:

Clear names, schema fields, and descriptions help the model choose and populate the intended operation; however, the critical mechanism here is that structured nested validation rejects incomplete or malformed records consistently, and tool schemas still need server-side validation of all inputs.

D:

Coordination or transactions prevent lost updates and inconsistent final state; this cannot substitute for the behavior that structured nested validation rejects incomplete or malformed records consistently, because independent model-generated reasoning does not provide concurrency control.

Question 379

A deployment agent must access a read-only dashboard but never issue write operations to production. What should be implemented to solve this problem?

  1. Use a local function dispatcher for private integrations
  2. Set a maximum tool iteration count
  3. Apply a per-run timeout
  4. Configure an allowlist of available tools

Explanation

Correct Answer

D:

Its registered and authorized tool set excludes mutating functions before Claude can request them; the implementation must still recognize that an instruction not to delete data is insufficient if destructive tools remain callable.

Incorrect Answers

A:

Application code executes a narrowly defined local tool and returns selected results to Claude; this cannot substitute for the behavior that its registered and authorized tool set excludes mutating functions before Claude can request them, because the API itself does not have arbitrary access to a customer's private internal network.

B:

A harness enforces a fixed ceiling on action cycles and reports incomplete execution safely; in contrast, its registered and authorized tool set excludes mutating functions before Claude can request them, while model self-assessment is too probabilistic to serve as the only loop brake.

C:

The host checks elapsed time and cancels or stops further tool iterations at the limit; this addresses a separate concern, not the fact that its registered and authorized tool set excludes mutating functions before Claude can request them; increasing token capacity does not impose a real execution deadline.

Question 380

Clients encounter throttling and the server communicates a suggested pause before retrying. Which action most directly closes the gap?

  1. Measure both request and token rates
  2. Respect retry-after guidance when available
  3. Use retry-aware HTTP error handling
  4. Queue overload rather than dropping accepted work

Explanation

Correct Answer

B:

Retry scheduling honors the provider's feedback and avoids synchronized hammering; the important constraint is that relying only on fixed zero-delay retries is counterproductive.

Incorrect Answers

A:

Capacity planning accounts for relevant request, input, and output token budgets, whereas the requirement calls for another behavior: retry scheduling honors the provider's feedback and avoids synchronized hammering; counting only completed HTTP calls can miss token-based constraints.

C:

Status classification drives different responses for transient and permanent errors; the requirement instead turns on this behavior: retry scheduling honors the provider's feedback and avoids synchronized hammering; blindly retrying invalid JSON does not improve correctness.

D:

A persistent queue buffers approved tasks with controlled worker concurrency; this cannot substitute for the behavior that retry scheduling honors the provider's feedback and avoids synchronized hammering, because unbounded in-process arrays risk memory exhaustion and data loss.

Claude Certified Developer - Foundations (CCDV-F) Practice Tests