Anthropic CCDV-F (Claude Certified Developer - Foundations) Practice Test 15

Question 140

Compliance analysts must verify where a reported liability cap appears in the uploaded agreement. Which development approach should be selected?

  1. Include source spans for extracted values
  2. Return an explicit unknown for missing evidence
  3. Separate citations from generated claims
  4. Escalate low-confidence high-impact cases

Explanation

Correct Answer

A:

Provenance fields or offsets let reviewers inspect the original supporting text, although a fluent explanation without a source is not independently verifiable.

Incorrect Answers

B:

A nullable or unknown state avoids inventing a plausible but unsupported date; however, the critical mechanism here is that provenance fields or offsets let reviewers inspect the original supporting text, and guessing a value would convert absence of evidence into false data.

C:

The application verifies source mapping and clearly displays provenance when available; the requirement instead turns on this behavior: provenance fields or offsets let reviewers inspect the original supporting text; formatting fake-looking references beautifully does not provide evidence.

D:

Risk-based human review catches consequential uncertain outputs before automatic action; this cannot substitute for the behavior that provenance fields or offsets let reviewers inspect the original supporting text, because automatic acceptance based on polished prose could cause compliance failures.

Question 141

A search-backed assistant often answers with sources unrelated to the claimed statements. Which change would satisfy the constraint?

  1. Monitor retrieval coverage and citation fidelity
  2. Use a golden evaluation set for regressions
  3. Treat retrieved pages as untrusted evidence
  4. Track side-effect authorization denials

Explanation

Correct Answer

A:

Measuring evidence relevance and attribution identifies grounding defects; notably, extra context tokens cannot guarantee faithful citation of the retrieved material.

Incorrect Answers

B:

Fixed labeled cases detect performance changes across relevant user segments; in contrast, measuring evidence relevance and attribution identifies grounding defects, while one successful demo query does not establish production quality.

C:

External page text should be handled as data, not as new operational authority; this addresses a separate concern, not the fact that measuring evidence relevance and attribution identifies grounding defects; a page author cannot legitimately override the application's system policy.

D:

Auditing denied calls helps identify policy confusion or prompt injection pressure; here, the controlling capability is that measuring evidence relevance and attribution identifies grounding defects, but denials should remain enforced even if error metrics worsen.

Question 142

A security-review procedure requires reusable guidance plus access to a repository metadata lookup. What is the correct engineering decision?

  1. Combine MCP resources with a summarization prompt
  2. Combine a Skill with a read-only tool
  3. Use a Skill for reusable procedures
  4. Combine a guarded custom tool with a human approval step

Explanation

Correct Answer

B:

The Skill supplies the workflow while the tool supplies authorized external facts, although instruction packages and executable integrations solve different needs.

Incorrect Answers

A:

Resources supply source data while a prompt governs how it should be analyzed; the needed capability is different—the Skill supplies the workflow while the tool supplies authorized external facts—and neither component independently guarantees the summary is factually faithful.

C:

Skills package instructional workflows that an agent can reuse within supported clients; however, the critical mechanism here is that the Skill supplies the workflow while the tool supplies authorized external facts, and they are guidance, not network APIs or execution permissions.

D:

The tool executes a constrained operation after an enforced business approval gate; the requirement instead turns on this behavior: the Skill supplies the workflow while the tool supplies authorized external facts; model-generated consent text is not an auditable approval event.

Question 143

A nightly import produces more inference jobs than the workers can process instantly. What change is best aligned with the goal?

  1. Use a queue for disconnected long-running workloads
  2. Queue overload rather than dropping accepted work
  3. Respect retry-after guidance when available
  4. Cap the number of concurrent requests

Explanation

Correct Answer

B:

A persistent queue buffers approved tasks with controlled worker concurrency—with this limitation: unbounded in-process arrays risk memory exhaustion and data loss.

Incorrect Answers

A:

Durable queued work survives client disconnection and exposes later status; however, the critical mechanism here is that a persistent queue buffers approved tasks with controlled worker concurrency, and keeping an open WebSocket alone does not make background jobs persistent.

C:

Retry scheduling honors the provider's feedback and avoids synchronized hammering; this cannot substitute for the behavior that a persistent queue buffers approved tasks with controlled worker concurrency, because relying only on fixed zero-delay retries is counterproductive.

D:

A concurrency pool smooths demand and limits rate-limit errors; in contrast, a persistent queue buffers approved tasks with controlled worker concurrency, while throughput capacity and latency vary by model and account limits.

Question 144

A machine client needs a fixed set of fields with well-defined types instead of arbitrary prose. Which response is most defensible under these conditions?

  1. Use named sections for long complex requirements
  2. Use explicit delimiters for source passages
  3. Use tool schemas for action interfaces
  4. Specify a strict response schema

Explanation

Correct Answer

D:

Schema constraints and subsequent validation improve consistent downstream parsing; notably, they cannot guarantee that the values are factually correct.

Incorrect Answers

A:

Structured headings make independent instructions easier to distinguish and check, whereas the requirement calls for another behavior: schema constraints and subsequent validation improve consistent downstream parsing; structure cannot rescue contradictory or factually impossible requirements.

B:

Marking the passage's bounds clarifies which content is data rather than a user command; however, the critical mechanism here is that schema constraints and subsequent validation improve consistent downstream parsing, and delimiters help interpretation but do not replace tool-level security.

C:

Tool definitions and server validation specify callable operations more reliably than prose parsing; the requirement instead turns on this behavior: schema constraints and subsequent validation improve consistent downstream parsing; a natural-language command is not an adequate machine-checked action contract.

Question 145

An assistant may search internal incident reports that include information restricted to another team. Which measure directly addresses the problem?

  1. Authorize each requested operation at the resource level
  2. Authenticate the caller before processing account actions
  3. Protect confidentiality during evidence retrieval
  4. Filter outgoing sensitive data at the egress layer

Explanation

Correct Answer

C:

Retrieval must filter documents by verified entitlement before adding snippets to context, subject to this requirement: asking Claude not to quote unauthorized documents does not prevent exposure upstream.

Incorrect Answers

A:

Access policies check whether that principal may perform the exact requested change on the target object; this cannot substitute for the behavior that retrieval must filter documents by verified entitlement before adding snippets to context, because successful login alone does not grant permission across all records.

B:

Identity verification establishes who is requesting the action before tool access is considered; in contrast, retrieval must filter documents by verified entitlement before adding snippets to context, while a convincing statement of ownership inside a prompt is not authentication.

D:

Deterministic output and action checks block unauthorized exfiltration before it leaves the trusted environment; here, the controlling capability is that retrieval must filter documents by verified entitlement before adding snippets to context, but model self-policing alone cannot guarantee detection of every malicious instruction.

Question 146

A service is rolling out a new prompt policy to only a subset of users. If misclassifications rise, engineers must revert it and identify precisely which release produced each trace. Which TWO configuration practices are needed? Choose TWO.

  1. Record prompt version in traces
  2. Maintain repository CLAUDE.md instructions
  3. Deploy prompts with feature flags
  4. Assign semantic versions to prompts
  5. Test a prompt change on a frozen dataset

Explanation

Correct Answers

A:

Trace metadata connects outcomes to the actual deployed instruction set; the important constraint is that a model name by itself does not explain prompt-specific behavior changes.

C:

Configurable release controls can switch a bounded population between approved prompt versions; notably, editing production text in place without history prevents controlled reversal.

Incorrect Answers

B:

Project memory files make scoped guidance available in Claude Code working contexts; the requirement instead turns on this behavior: trace metadata connects outcomes to the actual deployed instruction set; they should not contain credentials or private production secrets.

D:

Versioned templates link each inference to the rules actually active during its execution; in contrast, trace metadata connects outcomes to the actual deployed instruction set, while keeping only the newest text prevents reliable regression analysis.

E:

A regression suite compares outcomes with expected labels before promotion; this addresses a separate concern, not the fact that trace metadata connects outcomes to the actual deployed instruction set; a valid JSON response alone cannot establish behavior equivalence.

Question 147

A solution uses a specific reasoning configuration and tool set that might not be supported on every candidate model. Which development approach should be selected?

  1. Update cost and context assumptions
  2. Recheck behavior after version migration
  3. Verify tool and thinking feature support
  4. Check currently supported model IDs before deployment

Explanation

Correct Answer

C:

Capability checks prevent incompatible parameter and feature combinations; its operational boundary is that similar family names do not guarantee identical APIs or limits.

Incorrect Answers

A:

Recalculating budgets avoids relying on stale token or feature assumptions; the requirement instead turns on this behavior: capability checks prevent incompatible parameter and feature combinations; product names alone do not prove pricing or capacity parity.

B:

Regression tests evaluate prompts, tool arguments, latency, and output shape on the new version; this cannot substitute for the behavior that capability checks prevent incompatible parameter and feature combinations, because version changes can affect behavior even when the API request remains syntactically valid.

D:

The official model catalog identifies supported identifiers and feature availability; this addresses a separate concern, not the fact that capability checks prevent incompatible parameter and feature combinations; historical aliases should not be assumed to remain valid.

Question 148

A multi-tenant Messages API service needs reviewed customer-specific instruction revisions and immediate rollback without redeploying mobile clients. What is the most direct implementation choice?

  1. Use the Messages API system field for app-wide behavior
  2. Separate tenant-specific conversation state
  3. Use a user content block for the actual request
  4. Version tenant-specific instruction templates in application configuration

Explanation

Correct Answer

D:

Server-owned configuration versions the active prompt template for each tenant and permits controlled rollout or reversal; notably, versioning templates does not replace supplying trusted instructions in the API system field.

Incorrect Answers

A:

Server code supplies a governing system instruction separately from the user-message array, whereas the requirement calls for another behavior: server-owned configuration versions the active prompt template for each tenant and permits controlled rollout or reversal; placing it inside user-supplied text exposes the rule to unintended editing and trust confusion.

B:

Tenant isolation is enforced in storage queries, tools, and context assembly; however, the critical mechanism here is that server-owned configuration versions the active prompt template for each tenant and permits controlled rollout or reversal, and merely instructing the model not to mix tenants is an inadequate barrier.

C:

A user message carries the current request as task data subordinate to the system rules; the requirement instead turns on this behavior: server-owned configuration versions the active prompt template for each tenant and permits controlled rollout or reversal; treating customer input as a new system instruction would invert the intended trust hierarchy.

Question 149

Every submitted order must have its tax calculated by a fixed statutory formula. Which development approach should be selected?

  1. Use an Anthropic server-executed tool where supported
  2. Use an MCP-provided tool through a configured client
  3. Use a deterministic application function without model dispatch
  4. Validate and execute the tool outside the model

Explanation

Correct Answer

C:

Conventional code runs the calculation directly and avoids unnecessary probabilistic tool selection, although a language-model intermediary would add cost without changing the arithmetic rules.

Incorrect Answers

A:

Server tools execute on the service side and return associated results during the model interaction; by comparison, the intended solution relies on this: conventional code runs the calculation directly and avoids unnecessary probabilistic tool selection; they have distinct data and capability boundaries from the customer's custom functions.

B:

The MCP client discovers the server's tools and invokes permitted operations through a standardized protocol; the needed capability is different—conventional code runs the calculation directly and avoids unnecessary probabilistic tool selection—and MCP transport does not replace authentication or authorization of the business system.

D:

Application code enforces permission checks and performs the actual side effect or read; however, the critical mechanism here is that conventional code runs the calculation directly and avoids unnecessary probabilistic tool selection, and the presence of tool_use content is not evidence the backend was queried.

Question 150

An agent consumes attacker-controlled web pages and has access to protected tools. Which response is most defensible under these conditions?

  1. Perform security abuse tests
  2. Use regression evaluations for prompts
  3. Run unit tests for a tool argument validator
  4. Categorize and triage production incidents

Explanation

Correct Answer

A:

Red-team cases test injection resistance and authorization enforcement at trust boundaries, although ordinary happy-path tests will not expose deliberate instruction attacks.

Incorrect Answers

B:

A frozen dataset and grading rubric compare model outputs before and after the proposed prompt revision; however, the critical mechanism here is that red-team cases test injection resistance and authorization enforcement at trust boundaries, and passing a linter does not measure semantic output quality.

C:

Unit cases exercise the validator's accept/reject logic without live API traffic; the requirement instead turns on this behavior: red-team cases test injection resistance and authorization enforcement at trust boundaries; a complete agent test is slower and less precise for locating pure validation bugs.

D:

Incident classification routes problems to the correct service, integration, or prompt team; this cannot substitute for the behavior that red-team cases test injection resistance and authorization enforcement at trust boundaries, because treating every failure as bad model reasoning conceals infrastructure issues.

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