Question 60
One worker needs another's normalized data before analysis begins. How should dependency planning represent this?
- Repeat or replace only the failed subtask while keeping successful outputs
- Make data normalization a prerequisite for the dependent analysis
- Specify measurable deliverables and pass-fail conditions for each subtask
- Partition the issue into bounded states with explicit exit and escalation criteria
Explanation
Correct Answer
B:The task graph must express actual data dependencies. Running both phases in parallel can force the analyst to guess a schema or use stale inputs.
Incorrect Answers
A:Independence in the task graph permits selective replay; this contributes to isolating retry to failed work without meeting the need for reflecting real data dependencies in the task graph.
C:Because a task boundary is useful when the parent can decide whether its output satisfies requirements, the measure supports specifying task-level acceptance criteria rather than the needed reflecting real data dependencies in the task graph.
D:Bounded tasks define what evidence and action terminate work; its benefits for defining bounded task exits do not resolve the need for reflecting real data dependencies in the task graph.
Question 61
A support classifier mishandles cases with no matching policy category. Which few-shot example is most valuable?
- Add contrasting high-severity and low-severity labeled cases with rationales
- Check on held-out realistic tickets rather than judging only the shown demonstrations
- Add an example that correctly returns an unknown or escalation category
- Display diverse input date formats paired with one required normalized representation
Explanation
Correct Answer
C:Boundary examples teach that abstention is acceptable. Showing only easy positive categories encourages the model to force every novel case into a known class.
Incorrect Answers
A:Contrastive examples teach the decision boundary between categories; this may support distinguishing severity classes with contrastive examples but cannot alone meet the requirement for calibrating abstention with a boundary example.
B:In-prompt examples are teaching material, not an independent evaluation set; the stated problem requires calibrating abstention with a boundary example rather than checking generalization beyond few-shot examples.
D:For demonstrating cross-locale date normalization, contrasting examples demonstrate conversion rules while preserving a single target shape; however, this requirement concerns calibrating abstention with a boundary example.
Question 62
A design review claims that declaring a calculate_tax tool causes Claude to execute Python in the user database automatically. What correction is needed?
- Run the requested client tool and return a matching tool_result
- Check the side effect with an authoritative follow-up read before promising completion
- Keep the tool implementation in the authorized execution environment
- Check generated arguments against the declared tool input schema before running business logic
Explanation
Correct Answer
C:Client tools are interfaces advertised to the model; the hosting application runs their operations. A model-produced tool_use block is a request, not automatic privileged execution.
Incorrect Answers
A:A tool_use stop hands control to the application; it must run the requested operation and pass back a tool_result tied to the call identifier before Claude can finish reasoning with that evidence; that relates to completing the client-tool round trip, not the necessary distinguishing model intent from tool execution.
B:Although a tool's acknowledgement can be stale or incomplete relative to durable state, the choice addresses verifying externally visible side effects and not the requested distinguishing model intent from tool execution.
D:Tool schemas describe the callable interface, but application-side validation still protects execution paths and prevents malformed or unsafe arguments from reaching external systems; the option would serve checking tool input validity, not distinguishing model intent from tool execution as specified in the stem.
Question 63
A coordinator receives a worker claim with no indication that the worker inspected the target document. What handoff field helps?
- Require evidence of which documents and versions the worker actually consulted
- Maintain approval for destructive decisions with the coordinator or an authorized human
- Specify one shared input and output field contract before spawning workers
- Send condensed relevant findings and source handles rather than full unrelated transcripts
Explanation
Correct Answer
A:Source and version metadata distinguish inspected facts from memory-based guesses. The parent can then decide whether missing evidence warrants another tool call.
Incorrect Answers
B:When considering retaining authority for irreversible actions, delegating writing does not imply delegating business authority; the requirement for validating the evidentiary basis of a subagent answer remains unmet.
C:When addressing enforcing a shared worker data contract, consistent schemas make aggregation mechanical and reduce misbound identifiers; the present task instead calls for validating the evidentiary basis of a subagent answer.
D:Selecting task-relevant evidence controls token growth without erasing provenance; controlling delegated context size and validating the evidentiary basis of a subagent answer require different controls in this situation.
Question 64
Several workers independently inspect the same very large configuration folder. Which context strategy reduces duplication?
- Read package-scoped guidance while exploring the affected subtree
- Centralize a short verified configuration summary with exact file references
- Review the implicated method and immediate call chain before expanding scope
- Choose repository search to locate relevant symbols and then read focused files
Explanation
Correct Answer
B:Shared grounded notes prevent repeated expensive reads while retaining the ability to verify details. Workers can request only the relevant underlying files when needed.
Incorrect Answers
A:Local rules capture framework and testing differences across packages; using local conventions during monorepo exploration and reusing concise verified repository findings are distinct responsibilities under these constraints.
C:Error evidence supplies a high-signal starting point; this can serve using stack traces to prioritize code inspection, but leaves the requirement for reusing concise verified repository findings unmet.
D:Search-first exploration narrows the context to likely dependencies; that addresses narrowing large-repository inspection by search, whereas this case calls for reusing concise verified repository findings.
Question 65
An order lookup receives an HTTP timeout while the fulfillment system may have already processed the command. How should the error be classified?
- Unify structured error codes and retry hints across tools
- Mark the outcome as uncertain and request reconciliation before replay
- Separate a legitimate not-found result from a tool execution error
- Mask secrets in public tool errors and retain details in protected logs
Explanation
Correct Answer
B:A timeout does not prove nonexecution for mutating operations. Reporting an indeterminate result and reconciling state prevents blind duplicate side effects.
Incorrect Answers
A:Because consistent machine-readable categories make orchestration decisions repeatable, the measure supports normalizing tool error classifications rather than the needed handling an ambiguous post-request timeout.
C:A successful search with zero matches is not necessarily an infrastructure failure; supporting separating no-match outcomes from failures cannot replace the required handling an ambiguous post-request timeout.
D:Agent-visible diagnostics should be actionable without exposing credentials; this contributes to sanitizing error information returned to models without meeting the need for handling an ambiguous post-request timeout.
Question 66
A developer edits the MCP server configuration but Claude Code continues using an old connection. What operational step should be checked?
- Reconnect or reload the affected server and verify its active configuration
- Check required environment variables are injected in the CI runtime
- Inspect tool exposure, permission filters and server-side authorization
- Safeguard the transport with TLS and authenticate clients before sensitive tool calls
- Configure the integration in a local or user-controlled configuration scope
Explanation
Correct Answer
A:Running sessions may retain existing transport state. Confirming that a changed definition is actually active is more useful than assuming file edits instantly alter every connection.
Incorrect Answers
B:For supplying MCP configuration variables in CI, an MCP configuration that expands environment references depends on that process's environment; however, this requirement concerns applying modified MCP server settings.
C:Connectivity alone does not grant every capability; that may serve diagnosing a connected but unavailable MCP tool, but it is no substitute for applying modified MCP server settings.
D:Tool inputs and results may contain confidential operational data; this may support protecting an MCP network connection but cannot alone meet the requirement for applying modified MCP server settings.
E:Personal scope keeps experimental endpoints out of the shared project contract; the stated problem requires applying modified MCP server settings rather than keeping personal MCP configuration out of shared files.
Question 67
A document extractor needs multiple nested shipping addresses with constrained types. What should its schema express?
- Flag required properties and disallow additional properties in the output schema
- Establish the tool input schema with a numeric amount field and strict validation where supported
- Describe line items as an array with a defined item object schema
- Use nested object definitions with typed address properties
Explanation
Correct Answer
D:A nested schema preserves structural relationships and makes each field independently verifiable. Flattening everything into unconstrained text loses hierarchy and consistency.
Incorrect Answers
A:Required fields and closed object shapes prevent silently missing or invented keys; the deciding condition is constraining nested JSON extraction structures, whereas this action concerns forbidding missing and unexpected JSON fields.
B:Typed schema constraints differentiate numeric amounts from text; the option would serve enforcing typed arguments for a tool call, not constraining nested JSON extraction structures as specified in the stem.
C:The declared array type tells the consumer how to iterate through repeated records; this addresses constraining collection shape in extraction output, while the question is about constraining nested JSON extraction structures.
Question 68
An operations prompt says 'handle errors gracefully', but some tool failures must stop work. What change improves reliability?
- Describe the defect categories and require a cited code location for every finding
- Provide a clear output structure with explicit labels for evidence and risk rationale
- Name retryable failures, mandatory escalation conditions and forbidden fallbacks
- Establish redaction rules and permitted data fields in the expected summary
Explanation
Correct Answer
C:Different error classes require different branches. A generic graceful-degradation instruction can lead the model to invent data where it should instead halt.
Incorrect Answers
A:When addressing preventing unsupported PR review findings, evidence-based criteria distinguish observed defects from speculative criticism; the present task instead calls for defining precise failure-handling criteria.
B:Combining an evaluation rubric with an output contract makes responses comparable; standardizing the structure of risk assessments and defining precise failure-handling criteria require different controls in this situation.
D:Data minimization instructions make the output contract sensitive to privacy; in this scenario, minimizing sensitive data in summaries is secondary to defining precise failure-handling criteria.
Question 69
A regulator asks which tool outputs justified an automated decision made last week. What should the system have stored?
- Tag confirmed observations separately from estimates and unknowns
- Model the value as missing with a documented reason and prohibit invented identifiers
- Keep an auditable decision record linking tool observations, rules and approvals
- Handle retrieved document instructions as untrusted content rather than authority
- Insist on a matching source span or flag the clause as unsupported
- Capture retrieval time and a source version or approved snapshot reference
Explanation
Correct Answer
C:Reproducibility requires more than the final recommendation. Traceable observations and policy checks support later investigation of disputed automated outcomes.
Incorrect Answers
A:Provenance distinguishes system-of-record data from inference; this can serve separating verified status from estimates, but leaves the requirement for reconstructing the evidence behind past decisions unmet.
B:Confidence labels cannot make absent evidence real; that addresses documenting absence without manufacturing data, whereas this case calls for reconstructing the evidence behind past decisions.
D:External evidence may contain prompt injection; resisting prompt injection embedded in source material and reconstructing the evidence behind past decisions are distinct responsibilities under these constraints.
E:Evidence-grounded extraction ties fields to document locations; the action supports tracing extracted clauses to document spans but falls short of reconstructing the evidence behind past decisions in this case.
F:Live pages can change or disappear; this can serve preserving time-sensitive research provenance, but leaves the requirement for reconstructing the evidence behind past decisions unmet.
Question 70
A messages controller treats stop_sequence as proof a task was fully resolved. What is the sounder decision?
- Carry out the requested client tool and return a matching tool_result
- Treat the custom stopping sequence as a generation boundary and validate task completion separately
- Emit an explicit failed tool result with diagnostic information rather than fabricate data
- Preserve the tool implementation in the authorized execution environment
Explanation
Correct Answer
B:A stop_sequence signals an application-specified token sequence caused termination. It does not independently prove that business assertions, tool operations, or response schemas are complete.
Incorrect Answers
A:A tool_use stop hands control to the application; it must run the requested operation and pass back a tool_result tied to the call identifier before Claude can finish reasoning with that evidence; supporting completing the client-tool round trip cannot replace the required separating custom stop boundaries from success.
C:The tool-error signal should flow into the conversation so the agent can retry, escalate, or communicate limits; that serves surfacing client-tool execution failure, whereas the case requires separating custom stop boundaries from success.
D:Because client tools are interfaces advertised to the model; the hosting application runs their operations, the measure supports distinguishing model intent from tool execution rather than the needed separating custom stop boundaries from success.