Question 500
Users can request short data transformations and the assistant may propose executable scripts. How should this requirement be met?
- Execute generated code only in a restricted sandbox
- Parse generated JSON with a real parser
- Treat generated SQL as untrusted text
- Use parameterized queries for user values
Explanation
Correct Answer
A:Sandboxing limits filesystem, network, process, and credential exposure; notably, asking the model to promise safety is not an operating-system isolation mechanism.
Incorrect Answers
B:Standard JSON parsing catches syntactic errors instead of relying on visual plausibility; in contrast, sandboxing limits filesystem, network, process, and credential exposure, while successfully parsed JSON can still violate expected business types.
C:Query validation and read-only permissions constrain execution against approved tables and patterns; this addresses a separate concern, not the fact that sandboxing limits filesystem, network, process, and credential exposure; text generation does not certify the safety or accuracy of the SQL.
D:Parameters keep values separate from SQL structure and prevent injection from changing the query grammar; here, the controlling capability is that sandboxing limits filesystem, network, process, and credential exposure, but escaping output for HTML would not prevent SQL injection.
Question 501
A team is integrating tools with irreversible side effects and must clarify trust boundaries early. Which technical control is appropriate?
- Use automated integration tests before release
- Write a design review before implementation
- Conduct a staged production rollout
- Require pull-request review
Explanation
Correct Answer
B:Architecture review assesses interfaces, failure modes, and control ownership before code changes; in production, implementing immediately can lock in unsafe assumptions and costly redesign.
Incorrect Answers
A:Integration tests check how independently coded components actually interact under realistic conditions; the missing capability is more specific: architecture review assesses interfaces, failure modes, and control ownership before code changes; unit tests of isolated helpers cannot reveal all cross-component contract failures.
C:Canary releases compare a limited population against baseline before broad enablement; the needed capability is different—architecture review assesses interfaces, failure modes, and control ownership before code changes—and immediate full deployment magnifies regressions and complicates rollback.
D:Peer review examines diff, tests, and ownership before a controlled merge, whereas the requirement calls for another behavior: architecture review assesses interfaces, failure modes, and control ownership before code changes; direct unreviewed edits reduce accountability and increase regression risk.
Question 502
Operational policy requires a record of who changed an agent's permission settings and when. Which capability is necessary for this outcome?
- Audit configuration changes
- Validate configuration at startup
- Retain a tested rollback version
- Run a configuration-diff preflight
Explanation
Correct Answer
A:Change logs capture approvals, diffs, and rollout outcomes for accountable administration; notably, a chat transcript is not a complete record of production configuration edits.
Incorrect Answers
B:Early checks fail fast with a clear operator error before the service accepts work; in contrast, change logs capture approvals, diffs, and rollout outcomes for accountable administration, while deferring discovery to live traffic increases incident impact.
C:Keeping the previously tested artifact enables a measured rollback when safe; this addresses a separate concern, not the fact that change logs capture approvals, diffs, and rollout outcomes for accountable administration; a rollback should consider state migrations and outstanding jobs.
D:Automated checks compare effective values against the approved deployment baseline; here, the controlling capability is that change logs capture approvals, diffs, and rollout outcomes for accountable administration, but assuming staging and production are identical is not a verification method.
Question 503
An MCP client discovers multiple actions and needs to know each operation's purpose and parameter contract. Which of the following would work as specified?
- Advertise tool metadata clearly
- Return structured errors from tools
- Expose an MCP tool for an executable operation
- Version and test server interfaces
Explanation
Correct Answer
A:The server supplies discoverable names, descriptions, and input schemas for the tool set; its operational boundary is that misleading descriptions can cause inappropriate model tool selection.
Incorrect Answers
B:Explicit error signaling distinguishes tool failure from an empty successful result; this addresses a separate concern, not the fact that the server supplies discoverable names, descriptions, and input schemas for the tool set; swallowing errors can cause a model to hallucinate completion.
C:An MCP tool advertises callable functionality with input schemas and returns execution results; here, the controlling capability is that the server supplies discoverable names, descriptions, and input schemas for the tool set, but tool calls still need explicit runtime permissions and backend validation.
D:Compatibility tests and versioned rollout protect consumers from unexpected protocol or field changes; the missing capability is more specific: the server supplies discoverable names, descriptions, and input schemas for the tool set; unreviewed changes can break integrations even when the server starts successfully.
Question 504
A multi-turn agent must continue after a tool invocation without losing the operation Claude requested. Which of these alternatives serves the stated goal?
- Stop the loop on end_turn
- Track token and tool budgets per task
- Return a tool_result with the matching tool_use_id
- Preserve the assistant tool_use turn in history
Explanation
Correct Answer
D:The assistant tool request and the following user tool result remain in the ordered message history; the important constraint is that sending only the result removes context needed to interpret what was requested.
Incorrect Answers
A:The harness treats the completion reason as the signal to deliver the final answer and exit; by comparison, the intended solution relies on this: the assistant tool request and the following user tool result remain in the ordered message history; repeating requests after natural completion can create unnecessary tool cycles.
B:Metering links request usage and tool counts to the logical job or trace identifier; the needed capability is different—the assistant tool request and the following user tool result remain in the ordered message history—and a single final answer's length is not a complete measure of work performed.
C:The caller adds a user-side tool_result block tied to the earlier tool use identifier, whereas the requirement calls for another behavior: the assistant tool request and the following user tool result remain in the ordered message history; an unrelated message cannot reliably bind a result to the invocation it completes.
Question 505
A prompt change improves most examples but damages responses for a small group of rare customer intents. Which capability is necessary for this outcome?
- Monitor retrieval coverage and citation fidelity
- Use a golden evaluation set for regressions
- Use a user content block for the actual request
- Instrument output schema failure rate
Explanation
Correct Answer
B:Fixed labeled cases detect performance changes across relevant user segments; the implementation must still recognize that one successful demo query does not establish production quality.
Incorrect Answers
A:Measuring evidence relevance and attribution identifies grounding defects; this addresses a separate concern, not the fact that fixed labeled cases detect performance changes across relevant user segments; extra context tokens cannot guarantee faithful citation of the retrieved material.
C:A user message carries the current request as task data subordinate to the system rules; the missing capability is more specific: fixed labeled cases detect performance changes across relevant user segments; treating customer input as a new system instruction would invert the intended trust hierarchy.
D:Tracking validation failures isolates content contract problems from transport availability; by comparison, the intended solution relies on this: fixed labeled cases detect performance changes across relevant user segments; HTTP success is not equivalent to a valid business payload.
Question 506
Incident responders must determine why two otherwise identical requests produced different classifications. Which technical control is appropriate?
- Test a prompt change on a frozen dataset
- Assign semantic versions to prompts
- Record prompt version in traces
- Retain a tested rollback version
Explanation
Correct Answer
C:Trace metadata connects outcomes to the actual deployed instruction set; notably, a model name by itself does not explain prompt-specific behavior changes.
Incorrect Answers
A:A regression suite compares outcomes with expected labels before promotion; however, the critical mechanism here is that trace metadata connects outcomes to the actual deployed instruction set, and a valid JSON response alone cannot establish behavior equivalence.
B:Versioned templates link each inference to the rules actually active during its execution; the requirement instead turns on this behavior: trace metadata connects outcomes to the actual deployed instruction set; keeping only the newest text prevents reliable regression analysis.
D:Keeping the previously tested artifact enables a measured rollback when safe; in contrast, trace metadata connects outcomes to the actual deployed instruction set, while a rollback should consider state migrations and outstanding jobs.
Question 507
A translation must be reviewed for terminology after drafting, and the second step depends on the first text. Which production safeguard is most relevant?
- Prefer an agent for exploratory diagnostics
- Use an evaluator-optimizer loop
- Select a sequential prompt chain
- Apply orchestrator-worker decomposition
Explanation
Correct Answer
C:The first output becomes the next step's input in a deliberate dependency chain; a crucial qualification is that running these stages concurrently would deny the reviewer the draft it must inspect.
Incorrect Answers
A:An agent observes evidence and selects the next tool based on new findings; this addresses a separate concern, not the fact that the first output becomes the next step's input in a deliberate dependency chain; it still requires restricted read-only tools and a bounded investigation scope.
B:A grader produces targeted feedback that the writer consumes on the next bounded iteration; here, the controlling capability is that the first output becomes the next step's input in a deliberate dependency chain, but without a maximum attempts policy the system can burn tokens without improving the draft.
D:A coordinator decomposes the task and merges workers' results after their bounded subtasks complete; by comparison, the intended solution relies on this: the first output becomes the next step's input in a deliberate dependency chain; workers need nonoverlapping responsibilities and explicit result contracts to avoid duplication.
Question 508
A creative-writing service should diversify suggestions, an extraction job needs more consistent responses, and audit staff must understand that identical prompts cannot promise identical text. Which THREE model-generation practices match those requirements? Choose THREE.
- Use a fixed regression dataset for stable comparisons
- Benchmark thinking settings with task-specific tests
- Treat output as generated prediction, not a database read
- Use low-temperature settings when supported for consistency
- Avoid assuming byte-identical responses
- Use a more exploratory sampling setting when supported
Explanation
Correct Answers
D:Lower randomness can reduce sampling variability in supported configurations; in production, it does not guarantee deterministic output or eliminate model errors.
E:Generative sampling and platform execution can cause output variability, so compare validated properties instead; a hard textual hash is usually an unsuitable semantic acceptance metric.
F:Sampling parameters can trade predictability for variety in compatible model modes, although increased randomness may degrade factual precision and schema reliability.
Incorrect Answers
A:Repeated evaluation over the same representative cases reveals distributional quality differences; this addresses a separate concern, not the fact that lower randomness can reduce sampling variability in supported configurations; one identical answer does not prove full determinism across deployments.
B:Controlled quality and latency measurements show whether the chosen task benefits; here, the controlling capability is that lower randomness can reduce sampling variability in supported configurations, but assuming maximal thinking is universally optimal wastes resources.
C:The language model generates plausible text rather than querying a company's live inventory by itself; the missing capability is more specific: lower randomness can reduce sampling variability in supported configurations; it needs authorized retrieval to ground fast-changing private facts.
Question 509
Eight independent product descriptions should be analyzed concurrently and then merged into one summary. What is the most direct implementation choice?
- Use a model-directed agent loop
- Execute a parallel fan-out workflow
- Route requests with an LLM classifier
- Prefer a workflow for regulated approvals
Explanation
Correct Answer
B:Independent branches can run at once and a reducer collects their outputs, lowering wall-clock time; a crucial qualification is that parallel execution is unsafe if branches mutate the same resource without coordination.
Incorrect Answers
A:Claude selects an action after observing each result, with the application enforcing iteration and permission limits; here, the controlling capability is that independent branches can run at once and a reducer collects their outputs, lowering wall-clock time, but the autonomy adds variability and demands hard budgets, logs, and a stopping policy.
C:A routing call chooses among bounded destinations while downstream code owns the actual execution; by comparison, the intended solution relies on this: independent branches can run at once and a reducer collects their outputs, lowering wall-clock time; classification alone cannot search, repair, or complete an open-ended case.
D:Explicit branches and immutable records make the process stable under regulatory review; the needed capability is different—independent branches can run at once and a reducer collects their outputs, lowering wall-clock time—and free-form agent planning is unsuitable for mandatory stage-order guarantees.
Question 510
A synchronous title generator takes a supplied text snippet and needs no tool execution or planning. What is the correct engineering decision?
- Use a basic Messages API call
- Use the Claude Agent SDK
- Build a custom Messages API loop
- Stop the loop on end_turn
Explanation
Correct Answer
A:A single request produces the completion without a persistent orchestration loop, subject to this requirement: adding agent scaffolding increases latency for a task with no external actions.
Incorrect Answers
B:The SDK provides higher-level agent controls and integrations while the application supplies policies and oversight; here, the controlling capability is that a single request produces the completion without a persistent orchestration loop, but prebuilt orchestration does not remove responsibility for secrets and tool permission boundaries.
C:A developer-managed loop explicitly handles tool_use, executes tools, returns tool_result, and continues as needed; the missing capability is more specific: a single request produces the completion without a persistent orchestration loop; every error, retry, and termination branch must be implemented by the application.
D:The harness treats the completion reason as the signal to deliver the final answer and exit; by comparison, the intended solution relies on this: a single request produces the completion without a persistent orchestration loop; repeating requests after natural completion can create unnecessary tool cycles.