Question 40
A custom address-lookup tool fails because its upstream service times out. The harness has a pending tool_use identifier, and the lookup has not produced an address. How should the failure be returned to Claude?
- Return a matching tool_result with is_error=true and a concise, sanitized failure description.
- Put the entire exception, including credentials, into ordinary assistant text and omit the tool result.
- Generate a new unrelated tool-use identifier for the error.
- Return an empty address as a successful result.
Explanation
Correct Answer
A:The matching identifier connects the observation to the attempted lookup. An explicit error distinguishes failure from an empty successful result, allowing the application and model to follow their recovery policy without inventing an address or exposing sensitive diagnostics.
Incorrect Answers
B:That leaks sensitive information and bypasses the structured tool-response relationship. A sanitized error observation is sufficient; raw secrets are not needed for the model to understand that the lookup failed.
C:The result must refer to the pending invocation. A new identifier breaks that association and can leave the original tool call unresolved in the conversation.
D:An empty successful value can be interpreted as a real lookup outcome. It hides the difference between “no address exists” and “the service did not respond.”
Question 41
A request router must choose billing or technical support. A message plausibly belongs to both, and sending it to the wrong queue is costly. What should the routing design include?
- A longer generated explanation for either choice.
- An ambiguity path that requests clarification or escalates.
- A rule always selecting the first listed queue.
- Two independent final replies sent immediately.
Explanation
Correct Answer
B:Routing needs a defined outcome for insufficient evidence. A clarification path avoids forcing a confident category when the input does not support one.
Incorrect Answers
A:More prose does not supply missing evidence. An elaborate rationale can conceal uncertainty without improving the selected queue.
C:List position does not resolve semantic ambiguity. It creates a systematic routing bias unrelated to the customer’s actual issue.
D:Sending conflicting resolutions creates duplicate ownership. Parallel analysis may help, but the system still needs a coordinated routing decision.
Question 42
A workflow first identifies a document’s language, then selects a language-specific extraction template. Which scheduling arrangement preserves the dependency?
- Complete language identification before selecting the template.
- Choose the template from the final extraction result.
- Start both stages with identical initial inputs.
- Run template selection once per available language.
Explanation
Correct Answer
A:The second stage consumes the first stage’s result. Waiting for that result prevents selection from using an unknown or guessed language.
Incorrect Answers
B:That reverses the dependency: the extraction already needs the template. The proposed order cannot establish the required initial configuration.
C:The template stage then lacks its required classification. Concurrency cannot remove a real data dependency between the stages.
D:This performs unnecessary work and still requires selecting the correct outcome. It replaces a simple dependency with avoidable speculation.
Question 43
A draft needs independent spelling, factual-reference, and formatting checks. None changes the draft. What can reduce review latency without changing these checks?
- Let the first check to finish determine all results.
- Make each check wait for the previous check’s complete transcript.
- Run the checks concurrently and combine their findings.
- Run checks concurrently while each edits the shared draft.
Explanation
Correct Answer
C:Each reviewer reads the same fixed input without depending on another result. Parallel execution can shorten the combined wait while preserving separate findings.
Incorrect Answers
A:Completion speed does not establish coverage. Accepting only the fastest result discards the other required review dimensions.
B:That introduces dependencies absent from the task. Passing entire transcripts also adds context that the independent checks do not require.
D:Concurrent edits change the input during review. Findings may then refer to different document versions instead of one consistent draft.
Question 44
A migration request may involve an unknown number of components, discoverable only after repository inspection. Which arrangement fits this uncertainty?
- Require every worker to inspect and migrate the entire repository.
- Have an orchestrator inspect the task and assign bounded component work.
- Always create exactly two workers with predetermined component names.
- Skip inspection and select components from repository age.
Explanation
Correct Answer
B:The decomposition depends on discovered structure. An orchestrator can create appropriate assignments after inspection instead of assuming a fixed component count.
Incorrect Answers
A:Duplicating full responsibility wastes effort and risks competing changes. It does not exploit the component boundaries discovered during inspection.
C:A fixed decomposition may omit newly discovered components or assign nonexistent ones. The work structure must follow the repository evidence.
D:Age does not reveal the migration’s dependency structure. The assignments would rest on an unrelated proxy rather than actual affected components.
Question 45
A generation-and-review loop keeps alternating between two writing styles. Each review says only “improve it.” What is the most useful design correction?
- Treat every rewrite as improvement because it is newer.
- Send only the previous review score to the generator.
- Increase the number of reviewers without sharing criteria.
- Define an explicit rubric and a bounded revision policy.
Explanation
Correct Answer
D:A stable rubric tells the generator what counts as improvement. Revision limits prevent subjective oscillation from continuing without a justified stopping condition.
Incorrect Answers
A:Revision order does not establish quality. Automatically accepting the latest output ignores whether it satisfies the actual task requirements.
B:A score without actionable criteria does not explain what to change. The generator still lacks guidance for resolving the specific weaknesses.
C:More reviewers can introduce additional incompatible preferences. The missing element is a common acceptance standard, not simply more opinions.
Question 46
Three evaluators use the same flawed prompt and the same incorrect reference. All approve an answer. What does their agreement establish?
- The answer becomes correct after a majority vote.
- Agreement alone does not eliminate their shared error.
- Three approvals prove the answer independently correct.
- Only execution timing can explain the agreement.
Explanation
Correct Answer
B:The evaluations have a common failure source. Repetition can reproduce that mistake, so validation needs reliable evidence or meaningfully different checks.
Incorrect Answers
A:Voting aggregates opinions; it does not change the underlying facts. A majority cannot repair an incorrect reference used by every voter.
C:The checks are not independent merely because they ran three times. Shared assumptions can produce unanimous but mistaken judgments.
D:Identical flawed guidance is a direct alternative explanation. Timing is not evidence that the evaluations examined the answer correctly.
Question 47
A fixed-label classifier already meets quality and latency targets with one model call. No tools or multi-step decisions are required. What architecture should remain the baseline?
- A supervisor with four specialized agents by default.
- An autonomous loop that repeatedly reclassifies until unanimous.
- A tool-discovery layer for capabilities the task never uses.
- The measured single-call implementation.
Explanation
Correct Answer
D:Additional orchestration needs a demonstrated benefit. The existing implementation satisfies the requirements and provides a simpler baseline for maintenance and evaluation.
Incorrect Answers
A:More agents add coordination and inference overhead. The scenario provides no unmet requirement that requires those additional responsibilities.
B:Repeated classification creates extra cost and possible disagreement. It does not address a demonstrated failure in the accepted baseline.
C:Discovering unused capabilities adds complexity without improving classification. Architecture should follow required behavior rather than available abstractions.
Question 48
Two read-only search services can independently provide a complete answer. The workflow needs the earliest answer that passes a validator, rather than both results. Which orchestration rule fits?
- Wait for both services and require their answers to be identical.
- Accept the first validated result and stop unnecessary remaining work.
- Combine partial fields from both responses before validating either.
- Accept whichever service responds first, even if validation fails.
Explanation
Correct Answer
B:This is a first-valid-result race rather than a join of required contributions. A fast failed response must not win, and stopping the other read avoids work after the objective is satisfied.
Incorrect Answers
A:Agreement may be useful in another design, but this workflow explicitly accepts either independently valid answer. Waiting for both adds an unnecessary dependency.
C:Each service promises a complete result. Mixing unvalidated fields creates a third answer that neither service supplied and is not the requested race policy.
D:Completion order alone does not meet the validity requirement. An early error or malformed answer must leave the other search eligible to succeed.
Question 49
Several workers review independent packages, and package review times vary greatly. Fixed batches leave some workers idle while another has many packages remaining. Which scheduling design improves utilization without reviewing a package twice?
- Assign every pending package to every idle worker simultaneously.
- Wait until all workers finish before assigning any remaining package.
- Let idle workers atomically claim the next package from a shared queue.
- Keep fixed batches but let workers report progress more frequently.
Explanation
Correct Answer
C:Dynamic assignment lets available workers take unstarted work instead of waiting for their original batch boundaries. Atomic claiming preserves single ownership of each package while balancing variable durations.
Incorrect Answers
A:This can reduce idle time by duplicating effort, but it violates the requirement to avoid reviewing the same package twice.
B:A barrier preserves idle periods while a slow batch runs. Work can be assigned as capacity becomes available without waiting for every worker.
D:More frequent reporting exposes idle capacity but does not redistribute packages. Workers still wait while another retains its slower batch.
Question 50
A task contains thirty trivial string replacements, each cheaper to perform than to coordinate through another model call. Which approach is most efficient?
- Choose the largest model for each individual replacement.
- Create one specialist agent for every replacement.
- Ask each specialist to rediscover the replacement rule.
- Execute the replacements in ordinary code within one task.
Explanation
Correct Answer
D:Deterministic transformations need no separate reasoning agent. Local code avoids delegation overhead while producing directly testable edits.
Incorrect Answers
A:Model size does not justify repeated inference for a simple deterministic operation. The specified transformation can be implemented directly.
B:Coordination dominates the useful work in this scenario. Dividing the task further increases cost without adding meaningful reasoning capability.
C:The rule is already known. Re-deriving it independently wastes effort and creates opportunities for inconsistent interpretation.