Question 340
A team wants every review comment to point to a real code location. Which output requirement supports downstream publishing?
- Demand verifiable evidence before elevating the claim to a blocking finding
- Confine allowed tools and repository credentials to read-only review operations
- Require file path, line range and finding rationale in a validated result schema
- Focus on requesting machine-readable JSON output and validate its schema
Explanation
Correct Answer
C:Location-aware structured findings allow a bot to attach actionable comments. Narrative-only criticism lacks reliable anchors for automated PR feedback.
Incorrect Answers
A:Because model assertions are hypotheses until supported by inspected code or tests, the measure supports verifying CI review claims before blocking rather than the needed generating location-specific PR findings.
B:CI runners often have elevated tokens; its benefits for enforcing read-only CI review permissions do not resolve the need for generating location-specific PR findings.
D:Structured output is easier for CI to parse and gate deterministically; this contributes to parsing review output safely in CI without meeting the need for generating location-specific PR findings.
Question 341
A parent agent mixes speculative ideas and confirmed facts into the final answer after worker delegation. What should the merging stage enforce?
- Build in a deterministic aggregation policy for worker output acceptance
- Build in a targeted citation repair pass before further broad delegation
- Merge only evidence-tagged findings while retaining uncertainty annotations
- Escalate trivial deterministic steps directly and reserve specialists for nontrivial reasoning
Explanation
Correct Answer
C:Outputs from specialists are proposals until verified. A structured reconciliation pass can preserve claims, citations and confidence boundaries rather than laundering speculation into certainty.
Incorrect Answers
A:A merge contract can specify completeness, conflict handling and evidence precedence; a method suited to defining a reproducible worker merge policy does not meet the need for separating verified worker findings from guesses.
B:For targeted repair instead of repeated broad delegation, missing source attribution is a bounded post-processing defect; however, this requirement concerns separating verified worker findings from guesses.
D:Agent delegation has token, latency and observability costs; this may support avoiding needless specialist agent invocations but cannot alone meet the requirement for separating verified worker findings from guesses.
Question 342
A developer must update one API handler but does not know where routing is registered. What exploration strategy fits?
- Document tested commands, outcomes and affected paths alongside decisions
- Pinpoint references and consumer contracts before changing the shared definition
- Locate generation sources and edit the generator or schema instead of generated output
- Trace route registration and callers outward from the known endpoint name
Explanation
Correct Answer
D:A dependency-directed traversal identifies the local change impact. Random broad reading is less likely to reveal the exact handler chain efficiently.
Incorrect Answers
A:Implementation history is not reliable if verification evidence is lost; the deciding condition is tracing code paths from a known endpoint, whereas this action concerns preserving code-testing evidence in work summaries.
B:Shared APIs can affect distant packages; the option would serve mapping consumers of a shared interface, not tracing code paths from a known endpoint as specified in the stem.
C:Generated artifacts are often overwritten in the next build; this addresses distinguishing generated files from authoritative sources, while the question is about tracing code paths from a known endpoint.
Question 343
One tool call fails because the account lacks funds; another returns a rate-limit response with retry-after guidance. Which TWO distinct error-handling decisions are correct? Choose TWO.
- Schedule bounded retries after the server's rate-limit delay for the 429 response
- Label both insufficient funds and a rate-limit response as invalid-input errors and never retry either
- Retry the insufficient-funds operation with exponential backoff as though it were a temporary overload
- Mark the entire operation as fully successful whenever either of the two tool calls returns data
- Treat insufficient funds as a durable business-state error that needs a different next action
Explanation
Correct Answers
A:Honoring rate-limit backoff information adds a necessary safeguard: rate-limit responses can indicate when another attempt is reasonable; controlled backoff avoids synchronized retry storms and stays within service capacity.
E:Funds or policy failures generally need user or system-state changes, not immediate identical retries. A domain error should be distinct from an intermittent network fault.
Incorrect Answers
B:An insufficient-funds business constraint is nontransient, whereas a rate limit can clear after the server's indicated delay. Combining both under one permanent-error class mishandles 429s.
C:Time-based retries cannot create missing funds; they may amplify noise or repeated attempts. The account constraint needs an alternative business path, while only the rate limit merits bounded retry.
D:Success of one branch does not repair the other. The controller needs separate handling for the permanent business failure and the retryable 429 response.
Question 344
A developer says the task should be complete because the agent produced a summary, but unit tests were never run. What is missing?
- Escalate to identity clarification instead of advancing to account-specific tools
- Implement an executable policy gate before the final approval or refund call
- Record external-input and tool-pending states separately
- Focus on adding a test-pass acceptance condition before marking the coding task done
Explanation
Correct Answer
D:Stage completion is a business invariant that should be checked with tool output. A plausible textual summary is not evidence that code satisfies the acceptance tests.
Incorrect Answers
A:Identity-dependent steps require verified identifiers; this can serve blocking state transitions on missing identity, but leaves the requirement for using tests as a completion gate unmet.
B:A deterministic guard can compare amounts with configured limits; that addresses enforcing financial limits in a workflow, whereas this case calls for using tests as a completion gate.
C:Different waits have different timeouts, retries and authorization requirements; this concerns distinguishing human and tool handoff states, leaving the requirement for using tests as a completion gate outstanding.
Question 345
An analysis run has several independent hypotheses that need comparison. What session design supports fair evaluation?
- Document and compare policy versions before accepting the fork's recommendation
- Reconfirm the caller's authorization before returning to privileged tasks
- Store only necessary state and apply an explicit retention or deletion policy
- Fork from the same validated baseline and keep branch histories separate
Explanation
Correct Answer
D:A common starting context makes alternatives comparable. Separate histories prevent later assumptions from leaking into another hypothesis.
Incorrect Answers
A:A session branch can retain a different configuration snapshot; supporting detecting configuration drift across forks cannot replace the required comparing alternatives from a common session baseline.
B:A stored conversation is not a perpetual permission grant; this contributes to revalidating authorization on session resume without meeting the need for comparing alternatives from a common session baseline.
C:Resumability does not justify indefinite storage of sensitive context; that serves limiting persisted session data, whereas the case requires comparing alternatives from a common session baseline.
Question 346
A parent agent needs an independent critique without exposing its draft answer as an anchor. Which invocation strategy is appropriate?
- Narrow both delegated instructions and available data connectors to the assigned case
- Embed an explicit output contract and completion criteria in the delegated prompt
- Spawn a reviewer with an isolated prompt and the source requirements only
- Execute the two workers in separate contexts and compare at the parent level
Explanation
Correct Answer
C:A clean review context reduces anchoring on the first draft. The reviewer can judge against original evidence and instructions without inheriting the author's conclusions.
Incorrect Answers
A:Context minimization and access controls complement each other; that may serve narrowing a worker's data-access boundary, but it is no substitute for maintaining independence for critique.
B:A subagent needs a bounded objective, format and acceptance test; this may support making subagent completion testable but cannot alone meet the requirement for maintaining independence for critique.
D:Isolated contexts create genuinely independent assessments; a method suited to obtaining independent subagent hypotheses does not meet the need for maintaining independence for critique.
Question 347
An order-placement agent has submitted an inventory-reservation request, but the supplier has not confirmed any stock. The agent proposes to capture payment immediately. What state transition is appropriate?
- Remain in inventory-pending state until reservation succeeds, then allow payment capture.
- Capture payment immediately using a guessed reservation ID, and reconcile inventory later.
- Treat acceptance of the reservation request as confirmed stock before the operation completes.
- Request supervisor approval and bypass inventory confirmation to meet the response-time target.
Explanation
Correct Answer
A:State transitions must be driven by confirmed prerequisites, not model guesses. Charging before inventory is reserved risks payment without fulfillment; only a successful reservation should unlock the capture step.
Incorrect Answers
B:A guessed reservation identifier is not an acknowledgment. Advancing to payment on an unverified ID can charge for unavailable stock and leaves no authoritative evidence for reconciliation.
C:Request acceptance shows only that a reservation was submitted; it does not prove inventory was allocated. The state machine must await the outcome before it marks the order ready for payment.
D:Human approval might authorize an exceptional commercial action but does not create a stock reservation. The unsatisfied inventory prerequisite should still block the payment transition.
Question 348
A test set improves after adding examples, but nobody can identify which case helped. What evaluation practice supports tuning?
- Track example-set versions and compare outcomes on a fixed test suite
- Expose labeled cases where multiple classes are present and document the priority rule
- Curate examples representing distinct layouts and difficult ambiguity cases
- Exercise on held-out realistic tickets rather than judging only the shown demonstrations
Explanation
Correct Answer
A:Versioned few-shot experiments reveal the impact of changes. Without stable evaluation data, a new prompt may appear better because the tested examples changed.
Incorrect Answers
B:Mixed-category examples teach how conflicts resolve; in this scenario, teaching category precedence through examples is secondary to measuring few-shot changes with controlled evaluations.
C:When considering choosing representative examples rather than redundant ones, example diversity matters more than sheer count; the requirement for measuring few-shot changes with controlled evaluations remains unmet.
D:When addressing checking generalization beyond few-shot examples, in-prompt examples are teaching material, not an independent evaluation set; the present task instead calls for measuring few-shot changes with controlled evaluations.
Question 349
A monorepo wants coding conventions that every contributor gets when starting Claude Code in the project. Where should the team document shared instructions?
- Put repository-wide conventions only in one developer's personal instruction file
- Position additional directory-specific CLAUDE.md files near the relevant code
- Put reviewed repository conventions in the project-level CLAUDE.md
- Validate the intended CLAUDE.md is discovered and loaded for that working context
Explanation
Correct Answer
C:Project instruction files carry shared guidance alongside the codebase. Using a personal file would not ensure teammates see the same standards.
Incorrect Answers
A:Personal instructions are not reliably shared with every contributor. Reviewed project instructions should be stored in the repository-level CLAUDE.md.
B:Scoped instruction files can provide rules for particular subtrees; localizing instructions to code directories and distributing shared repository coding instructions are distinct responsibilities under these constraints.
D:Rules are useful only when the assistant encounters the relevant file and scope; this can serve verifying instruction-file loading, but leaves the requirement for distributing shared repository coding instructions unmet.
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.