Question 350
A generator always marks its own confident outputs as correct. What evaluation structure improves credibility?
- Grade findings by impact and evidence strength before producing recommendations
- Keep versioned prompts and rubrics and rerun a fixed regression evaluation suite
- Run an independent verifier against the original requirements and observed tool evidence
- Prepare a labeled test set including true positives, negatives and borderline cases
Explanation
Correct Answer
C:Self-assessment tends to reuse the model's assumptions. A separate validation path can detect missing acceptance criteria and unsupported completion claims.
Incorrect Answers
A:Review output should prioritize well-supported important issues; that serves prioritizing findings after independent review, whereas the case requires reducing self-confirmation bias in generated artifacts.
B:Because comparable benchmarks reveal behavior changes when the system evolves, the measure supports making multipass quality evaluations reproducible rather than the needed reducing self-confirmation bias in generated artifacts.
D:Calibration requires examples with known outcomes, not only successful demonstrations; supporting measuring review precision against ground truth cannot replace the required reducing self-confirmation bias in generated artifacts.
Question 351
Claude Code discovers an unfamiliar project MCP server definition after checking out a new repository. What is the appropriate trust practice?
- Review and approve the project server before granting it access
- Begin by deriving or verify tenant authorization on the server for every tool request
- Use MCP tool discovery and declared tool interfaces as the basis for decisions
- Guard the transport with TLS and authenticate clients before sensitive tool calls
Explanation
Correct Answer
A:Repository-supplied configurations can execute processes or contact services. Workspace trust and server approval help prevent unreviewed instructions from running with developer privileges.
Incorrect Answers
B:User-supplied identifiers are untrusted inputs, not authorization; a method suited to enforcing tenant isolation at the MCP backend does not meet the need for reviewing untrusted project MCP definitions.
C:For discovering MCP tools through protocol metadata, mCP presents structured capabilities with names, descriptions and schemas; however, this requirement concerns reviewing untrusted project MCP definitions.
D:Tool inputs and results may contain confidential operational data; that may serve protecting an MCP network connection, but it is no substitute for reviewing untrusted project MCP definitions.
Question 352
The agent found the exact test failure and a one-line condition fixes it. What iteration style is efficient?
- Make the minimal fix and rerun the focused failing test
- Inspect dependencies and produce a change plan before mutation
- Maintain a read-only mode throughout the assessment
- Locate specific tests and acceptance criteria alongside the edit steps
Explanation
Correct Answer
A:When root cause and scope are known, direct implementation with verification reduces churn. A broad rewrite would introduce unnecessary regression risk.
Incorrect Answers
B:Discovery identifies relevant modules, tests and side effects; the deciding condition is applying and checking a known minimal fix, whereas this action concerns mapping dependencies before modifying unfamiliar code.
C:An assessment boundary prohibits side effects regardless of how confident the proposed fix appears; the option would serve honoring a read-only operating constraint, not applying and checking a known minimal fix as specified in the stem.
D:A useful plan describes not just changes but the evidence that will establish success; this addresses embedding verification into the implementation plan, while the question is about applying and checking a known minimal fix.
Question 353
A compliance summary must not start until the evidence reviewer signs off on sources. What ordering should the agent graph use?
- Restrict delegation depth and concurrent subagent count at the orchestration layer
- Implement a deterministic aggregation policy for worker output acceptance
- Delegate analysis-only capabilities to the reviewer and editing tools to a separate implementer
- Make synthesis depend on the completed evidence-review stage
Explanation
Correct Answer
D:A dependency edge enforces that the reporting worker receives validated source material. Starting synthesis at the same time risks producing claims before verification is available.
Incorrect Answers
A:When considering constraining recursive multi-agent delegation, limits on nesting and worker breadth bound cost and latency; the requirement for enforcing a dependency before synthesis remains unmet.
B:When addressing defining a reproducible worker merge policy, a merge contract can specify completeness, conflict handling and evidence precedence; the present task instead calls for enforcing a dependency before synthesis.
C:Tool partitioning enforces role boundaries; separating creator and reviewer privileges and enforcing a dependency before synthesis require different controls in this situation.
Question 354
An OCR model supplies a low-quality postal address that cannot be validated. What quality gate should apply?
- Obtain a better source or human review instead of repeated identical retries
- Remove duplicated artifacts from extracted records using source identifiers and document context
- Verify totals and reject or review arithmetic inconsistencies
- Send uncertain critical addresses to a manual exception review
Explanation
Correct Answer
D:For shipping or compliance, uncertain addresses have operational consequences. A human gate prevents unverified location data from flowing into fulfillment.
Incorrect Answers
A:An absent source page cannot be recovered by sampling the same incomplete input; the action supports recognizing a nonrecoverable extraction input gap but falls short of routing low-confidence addresses to human review in this case.
B:Overlapping page captures can lead to double counting; this can serve removing duplicate extraction artifacts, but leaves the requirement for routing low-confidence addresses to human review unmet.
C:Semantic checks compare extracted values to business invariants; that addresses checking arithmetic consistency in extracted records, whereas this case calls for routing low-confidence addresses to human review.
Question 355
A research report must separate supported findings from speculation. What should the prompt require?
- Specify retryable failures, mandatory escalation conditions and forbidden fallbacks
- Demand source-backed claims and a labeled unknowns or uncertainty section
- Declare required positive findings, limitations, risks and outstanding decisions
- Document redaction rules and permitted data fields in the expected summary
Explanation
Correct Answer
B:Requiring evidence mapping and explicit uncertainty discourages treating unverified inference as fact. The reader can then distinguish grounded conclusions from gaps.
Incorrect Answers
A:Different error classes require different branches; its benefits for defining precise failure-handling criteria do not resolve the need for separating sourced conclusions from unknowns.
C:Balanced section requirements make omissions more visible during validation; this contributes to eliciting both benefits and risks in reports without meeting the need for separating sourced conclusions from unknowns.
D:Data minimization instructions make the output contract sensitive to privacy; that serves minimizing sensitive data in summaries, whereas the case requires separating sourced conclusions from unknowns.
Question 356
A chatbot is asked to make a high-value refund outside its delegated authority. What should its response convey?
- Save the exception rationale and route it through the authorized review process
- State the approval dependency and avoid claiming the refund has occurred
- Obtain policy-owner resolution before applying either disputed instruction
- Obtain authorized identity clarification before account-specific action
Explanation
Correct Answer
B:Transparent status respects the user's need for clarity without misrepresenting incomplete action. Authorized humans must decide the exception before payment execution.
Incorrect Answers
A:For processing policy exceptions through approval, unusual circumstances may justify discretion but need a human or policy-sanctioned decision; however, this requirement concerns communicating a pending human approval accurately.
C:Conflicting authoritative sources create genuine uncertainty; this may support escalating conflicting governing policies but cannot alone meet the requirement for communicating a pending human approval accurately.
D:Missing identity is a safety-critical ambiguity; the stated problem requires communicating a pending human approval accurately rather than escalating unresolved customer identity.
Question 357
Several MCP tools return different ad hoc error phrases for the same rate-limit condition. Which design makes recovery consistent?
- Standardize structured error codes and retry hints across tools
- Strip secrets in public tool errors and retain details in protected logs
- Provide a safe error category and actionable next-step detail
- Focus on reexecuting time-sensitive reads after recovery and replace stale failure state
Explanation
Correct Answer
A:Consistent machine-readable categories make orchestration decisions repeatable. Free-form strings require brittle text parsing and can cause conflicting retry policies.
Incorrect Answers
B:Agent-visible diagnostics should be actionable without exposing credentials; this addresses sanitizing error information returned to models, while the question is about normalizing tool error classifications.
C:Structured failure context allows the caller to distinguish bad inputs, authentication problems and outages; that relates to making tool failure diagnostic rather than opaque, not the necessary normalizing tool error classifications.
D:Although cached failure artifacts are not live evidence, the choice addresses refreshing state after a transient service failure and not the requested normalizing tool error classifications.
Question 358
A testing skill asks Claude to install new packages silently before any unit test. How should it be revised?
- Define a project skill containing the ADR template and review criteria
- Prepare the release workflow as a project skill with clear instructions
- Document required headings, evidence fields and a completion checklist
- Separate dependency installation behind explicit approval and documented prerequisites
Explanation
Correct Answer
D:Skills are reusable instructions, not a blanket permission grant for arbitrary software installation. Sensitive environment changes should remain governed by tool policy.
Incorrect Answers
A:A skill can provide consistent document structure and decision checkpoints; reusing a controlled architecture-record template and requiring approval for skill-driven dependency changes require different controls in this situation.
B:A reusable skill can collect steps, supporting files and tool permissions under one invocable workflow; the unsatisfied requirement is requiring approval for skill-driven dependency changes, not making a repeatable release workflow reusable.
C:A bounded output contract makes a reusable workflow easier to verify; in this scenario, standardizing a command's output shape is secondary to requiring approval for skill-driven dependency changes.
Question 359
A research assistant receives stop_reason end_turn after synthesizing cited findings. Which loop action is appropriate?
- Link each tool_result with the actual tool_use_id that created it
- Align and execute both tool calls, returning their respective results
- Return the completed assistant message to the caller
- Dispatch the result back into the model loop so Claude can decide the next action
Explanation
Correct Answer
C:The end_turn reason indicates the model finished its turn; the orchestration loop can surface the produced answer rather than manufacturing another tool step.
Incorrect Answers
A:Tool-result correlation is by the invocation identifier, not merely by a shared tool name; that addresses correlating tool responses by invocation, whereas this case calls for recognizing a completed model turn.
B:A single model response can request multiple client tools; the host should process each authorized call and retain its unique tool-use identifier so results cannot be crossed; this concerns handling parallel client-tool calls, leaving the requirement for recognizing a completed model turn outstanding.
D:The application does not simply display raw tool output as the final assistant response; the action supports resuming model reasoning after a result but falls short of recognizing a completed model turn in this case.
Question 360
A hook script itself crashes during a mandatory policy check. What production behavior is safest?
- Check protected-path checks in a pre-tool permission handler
- Conduct a PostToolUse handler after successful file-edit tools
- Read result payload semantics in a post-tool validation hook
- Fail closed or suspend the protected action according to explicit policy
Explanation
Correct Answer
D:When the decision service is unavailable, a high-risk action should not silently proceed. The policy needs a defined error path and observable failure state.
Incorrect Answers
A:A programmatic check can compare the requested path to restricted locations before access occurs; this contributes to denying sensitive file reads before access without meeting the need for handling a failed mandatory policy check.
B:A post-tool callback can inspect the completed edit and trigger formatting or verification in response; that serves reacting to a completed editing tool, whereas the case requires handling a failed mandatory policy check.
C:Because transport success does not necessarily mean business success, the measure supports detecting business failure despite transport success rather than the needed handling a failed mandatory policy check.