Anthropic CCDV-F (Claude Certified Developer - Foundations) Practice Test 13

Question 120

An incident investigator uses vendor documents, some outdated and some containing embedded instructions to the assistant. Which TWO context-handling measures keep the documents useful as evidence without adopting their directives or stale claims? Choose TWO.

  1. Label external text as untrusted input
  2. Preserve immutable business facts in structured state
  3. Attach source IDs to retrieved excerpts
  4. Track source timestamps or versions
  5. Keep unresolved questions explicit

Explanation

Correct Answers

A:

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

D:

Version metadata allows the agent to prefer current evidence and flag historical material—with this limitation: timestamp alone does not establish which source is authoritative.

Incorrect Answers

B:

The application records authoritative facts outside model-generated summaries and supplies them when needed, whereas the requirement calls for another behavior: trust-boundary marking prevents source content from being promoted to application authority; conversation recollection can drift or omit earlier commitments.

C:

Citation metadata lets downstream logic connect statements to original retrieved material; however, the critical mechanism here is that trust-boundary marking prevents source content from being promoted to application authority, and without accurate source mapping, fabricated references can look convincing.

E:

Recording open issues prevents compaction from converting uncertainty into an invented guarantee; this cannot substitute for the behavior that trust-boundary marking prevents source content from being promoted to application authority, because suppressing gaps for a cleaner paragraph undermines factual reliability.

Question 121

A long investigation approaches the context limit but must preserve key conclusions and pending questions. Which option best fits the described workflow?

  1. Summarize old steps before continuing
  2. Start a fresh isolated context for a subtask
  3. Write durable memory only through an authorized tool
  4. Use an external workflow engine

Explanation

Correct Answer

A:

A compact summary can replace bulky intermediate transcripts for the next model call; the implementation must still recognize that summaries can lose details unless critical facts and provenance are retained.

Incorrect Answers

B:

A separate worker receives only the necessary files and reports a minimal result; the missing capability is more specific: a compact summary can replace bulky intermediate transcripts for the next model call; isolation only works when tool rights and returned text follow the same boundary.

C:

A storage tool persists a validated fact in an application-owned record; by comparison, the intended solution relies on this: a compact summary can replace bulky intermediate transcripts for the next model call; model response text alone does not update a durable external database.

D:

A job engine coordinates state transitions and recovery outside the model's transient response; the needed capability is different—a compact summary can replace bulky intermediate transcripts for the next model call—and it still requires model calls as activities rather than relying on prompts for transaction control.

Question 122

Product owners need to know whether slowness comes from retrieval, Claude generation, or an external tool. What change is best aligned with the goal?

  1. Use a sandbox backend for tool integration tests
  2. Count errors by class and endpoint
  3. Record latency per integration stage
  4. Correlate downstream calls with request IDs

Explanation

Correct Answer

C:

Timed spans separate request phases and identify the bottleneck for targeted optimization; total elapsed time alone cannot localize the slow dependency.

Incorrect Answers

A:

A nonproduction environment exercises authorization, payload mapping, and failures safely; the missing capability is more specific: timed spans separate request phases and identify the bottleneck for targeted optimization; faking every backend call can hide actual integration incompatibilities.

B:

Structured metrics reveal whether the issue is client validation, quota pressure, or upstream availability; by comparison, the intended solution relies on this: timed spans separate request phases and identify the bottleneck for targeted optimization; a single error counter masks very different remediation paths.

D:

Propagated trace identifiers let operators reconstruct a request's causal path, whereas the requirement calls for another behavior: timed spans separate request phases and identify the bottleneck for targeted optimization; comparing timestamps manually is unreliable under concurrency.

Question 123

A wording change might alter customer categorization even if the API integration code is untouched. What is the most suitable pattern?

  1. Deploy prompts with feature flags
  2. Audit configuration changes
  3. Record prompt version in traces
  4. Test a prompt change on a frozen dataset

Explanation

Correct Answer

D:

A regression suite compares outcomes with expected labels before promotion; the implementation must still recognize that a valid JSON response alone cannot establish behavior equivalence.

Incorrect Answers

A:

Configurable release controls can switch a bounded population between approved prompt versions; this cannot substitute for the behavior that a regression suite compares outcomes with expected labels before promotion, because editing production text in place without history prevents controlled reversal.

B:

Change logs capture approvals, diffs, and rollout outcomes for accountable administration; in contrast, a regression suite compares outcomes with expected labels before promotion, while a chat transcript is not a complete record of production configuration edits.

C:

Trace metadata connects outcomes to the actual deployed instruction set; this addresses a separate concern, not the fact that a regression suite compares outcomes with expected labels before promotion; a model name by itself does not explain prompt-specific behavior changes.

Question 124

A new policy prompt needs limited rollout and rapid rollback if it increases refusals. Which of the following would work as specified?

  1. Review plugin permissions before activation
  2. Record prompt version in traces
  3. Deploy prompts with feature flags
  4. Assign semantic versions to prompts

Explanation

Correct Answer

C:

Configurable release controls can switch a bounded population between approved prompt versions; its operational boundary is that editing production text in place without history prevents controlled reversal.

Incorrect Answers

A:

Least-privilege permission review limits what installed integrations can access; the requirement instead turns on this behavior: configurable release controls can switch a bounded population between approved prompt versions; a friendly plugin description is not a security audit.

B:

Trace metadata connects outcomes to the actual deployed instruction set; this cannot substitute for the behavior that configurable release controls can switch a bounded population between approved prompt versions, because a model name by itself does not explain prompt-specific behavior changes.

D:

Versioned templates link each inference to the rules actually active during its execution; this addresses a separate concern, not the fact that configurable release controls can switch a bounded population between approved prompt versions; keeping only the newest text prevents reliable regression analysis.

Question 125

A team chooses between two Claude variants for long financial document extraction. Which mechanism should the team prioritize?

  1. Refresh prompt evaluations after upgrade
  2. Recheck behavior after version migration
  3. Verify tool and thinking feature support
  4. Benchmark quality on the actual task

Explanation

Correct Answer

D:

A representative evaluation suite reveals which model handles the real format and edge cases; the implementation must still recognize that a marketing ranking cannot replace workload-specific measurement.

Incorrect Answers

A:

Re-running the fixed evaluation set detects behavior shifts requiring prompt adjustment; this cannot substitute for the behavior that a representative evaluation suite reveals which model handles the real format and edge cases, because copying old pass rates forward would misstate new performance.

B:

Regression tests evaluate prompts, tool arguments, latency, and output shape on the new version; in contrast, a representative evaluation suite reveals which model handles the real format and edge cases, while version changes can affect behavior even when the API request remains syntactically valid.

C:

Capability checks prevent incompatible parameter and feature combinations; this addresses a separate concern, not the fact that a representative evaluation suite reveals which model handles the real format and edge cases; similar family names do not guarantee identical APIs or limits.

Question 126

An agent has finished exploration but needs space for final synthesis of verified findings. What should be implemented to solve this problem?

  1. Keep pointers to source documents
  2. Keep unresolved questions explicit
  3. Summarize intermediate reasoning and preserve outcomes
  4. Retain exact quoted requirements when necessary

Explanation

Correct Answer

C:

Compaction keeps conclusions, evidence, and open questions while discarding low-value intermediate tokens; notably, overaggressive summarization can remove the details that make conclusions defensible.

Incorrect Answers

A:

References and identifiers enable later retrieval without copying the entire source into every prompt; however, the critical mechanism here is that compaction keeps conclusions, evidence, and open questions while discarding low-value intermediate tokens, and pointers alone do not make an inaccessible document available at inference time.

B:

Recording open issues prevents compaction from converting uncertainty into an invented guarantee; the requirement instead turns on this behavior: compaction keeps conclusions, evidence, and open questions while discarding low-value intermediate tokens; suppressing gaps for a cleaner paragraph undermines factual reliability.

D:

Exact critical terms should remain intact as source data rather than be loosely paraphrased; in contrast, compaction keeps conclusions, evidence, and open questions while discarding low-value intermediate tokens, while a small numerical wording change can reverse a regulatory decision.

Question 127

Multiple customers share an application and each may read only its own saved discussions. Which implementation mechanism is needed?

  1. Authorize access to stored conversation history
  2. Redact secrets from observability logs
  3. Authorize each requested operation at the resource level
  4. Apply retention and deletion rules to transcripts

Explanation

Correct Answer

A:

Access-control checks at the storage layer prevent cross-account conversation disclosure; the important constraint is that a hidden session identifier is not enough to authorize retrieval.

Incorrect Answers

B:

Redaction removes sensitive fields before events reach metrics and log sinks; the requirement instead turns on this behavior: access-control checks at the storage layer prevent cross-account conversation disclosure; encrypting logs does not justify logging unnecessary secrets indefinitely.

C:

Access policies check whether that principal may perform the exact requested change on the target object; this cannot substitute for the behavior that access-control checks at the storage layer prevent cross-account conversation disclosure, because successful login alone does not grant permission across all records.

D:

Storage controls govern how long records, logs, and derived artifacts remain available; in contrast, access-control checks at the storage layer prevent cross-account conversation disclosure, while a model's context window does not define legal data retention requirements.

Question 128

A tool can return thousands of records but the agent requires only one account's approved status. Which engineering practice is the strongest match?

  1. Use targeted tool requests instead of bulk dumps
  2. Remove stale tool outputs
  3. Put stable instructions ahead of volatile user material
  4. Filter irrelevant document chunks

Explanation

Correct Answer

A:

Narrow queries reduce context consumption and sensitive data exposure simultaneously—with this limitation: returned results still need authorization before the query executes.

Incorrect Answers

B:

Pruning outdated results frees context for more relevant evidence and current questions; this cannot substitute for the behavior that narrow queries reduce context consumption and sensitive data exposure simultaneously, because removing the sole authoritative citation would weaken the final answer.

C:

A stable prefix supports prompt-cache reuse and clearer separation of policy from per-request data; in contrast, narrow queries reduce context consumption and sensitive data exposure simultaneously, while if user text appears before the checkpoint it can invalidate shared cache hits.

D:

Relevance reranking can reduce noise while keeping evidence that supports the answer; this addresses a separate concern, not the fact that narrow queries reduce context consumption and sensitive data exposure simultaneously; excessive filtering can discard essential exceptions or contradictory passages.

Question 129

A team optimizes extraction precision while preserving recall for a regulated document workflow. What is the most suitable pattern?

  1. Construct a representative validation set
  2. Put customer questions in user messages
  3. Retain previously passing edge cases
  4. Compare revisions on fixed metrics

Explanation

Correct Answer

D:

Measured pass rates and error types reveal tradeoffs between candidates, although choosing solely by stylistic preference can hide data-quality regressions.

Incorrect Answers

A:

A labeled suite of normal, edge, and adversarial examples measures practical performance; the missing capability is more specific: measured pass rates and error types reveal tradeoffs between candidates; one anecdotal success does not establish production reliability.

B:

The user role carries the requested task without altering the registered tool definitions; by comparison, the intended solution relies on this: measured pass rates and error types reveal tradeoffs between candidates; making every customer turn a system rule would collapse trust separation.

C:

Regression fixtures preserve known correct behavior across prompt changes; the needed capability is different—measured pass rates and error types reveal tradeoffs between candidates—and evaluating only newly fixed failures gives a biased view of quality.

Question 130

A Python agent needs typed inputs, validated outputs, and dependency injection into tools. What is the most suitable pattern?

  1. Use a fallback to a human decision-maker
  2. Use Strands-style model-driven agents
  3. Use typed PydanticAI-style dependencies
  4. Use LangGraph-style stateful orchestration

Explanation

Correct Answer

C:

A typed framework aligns model interactions with application data models and testable dependencies; the important constraint is that types validate shape, not the truth of every generated factual assertion.

Incorrect Answers

A:

Escalation transfers the ambiguous decision with relevant evidence and uncertainties; the needed capability is different—a typed framework aligns model interactions with application data models and testable dependencies—and generating a confident answer is not an acceptable replacement for the required human approval.

B:

An agent framework coordinates tool selection while exposing configuration and observability hooks, whereas the requirement calls for another behavior: a typed framework aligns model interactions with application data models and testable dependencies; tool permission and termination policies remain the integrator's responsibility.

D:

A graph framework makes transitions and persisted workflow state first-class constructs; the requirement instead turns on this behavior: a typed framework aligns model interactions with application data models and testable dependencies; the framework does not eliminate the need to secure the nodes' tools.

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