Question 200
A product owner describes an outcome but cannot decide between two incompatible data models. What should the agent do first?
- Examine references and present a removal plan before deletion
- Surface the architectural trade-off and request the missing business decision
- Obtain performance evidence and define the target metric before code changes
- Avoid moving from inspection to write-enabled execution
Explanation
Correct Answer
B:An unresolved design choice is not safe to infer from code alone. Planning should expose the decision to an authorized stakeholder before irreversible implementation.
Incorrect Answers
A:Deleting files on a hunch can break hidden consumers; that serves researching destructive refactors before execution, whereas the case requires resolving architecture ambiguity before editing.
C:Optimization without a baseline makes success impossible to judge and may harm unrelated paths; its benefits for baselining performance before optimization do not resolve the need for resolving architecture ambiguity before editing.
D:An assessment boundary prohibits side effects regardless of how confident the proposed fix appears; supporting honoring a read-only operating constraint cannot replace the required resolving architecture ambiguity before editing.
Question 201
CI runs Claude on untrusted pull-request content using a broad deployment token. What is the critical defense?
- Establish explicit severity criteria and require evidence for blocking findings
- Remove secrets and keep tokens in the runner's protected secret store
- Use least-privilege credentials and sandbox the untrusted review workload
- Mandate verifiable evidence before elevating the claim to a blocking finding
Explanation
Correct Answer
C:PR diffs can contain prompt injection or malicious scripts. Restricted secrets and execution environments stop untrusted content from inheriting deployment authority.
Incorrect Answers
A:Reviewer outputs should distinguish verified blockers from optional recommendations; the stated problem requires limiting trust for untrusted pull-request code rather than calibrating automated code-review severity.
B:Build logs are often retained and widely visible; a method suited to keeping CI credentials out of generated logs does not meet the need for limiting trust for untrusted pull-request code.
D:Model assertions are hypotheses until supported by inspected code or tests; that may serve verifying CI review claims before blocking, but it is no substitute for limiting trust for untrusted pull-request code.
Question 202
A downstream system requires an exact currency amount but the image is partly obscured. What output rule is appropriate?
- Recheck totals and reject or review arithmetic inconsistencies
- Attempt again with targeted feedback identifying the invalid field and expected format
- Abstain or mark the amount uncertain pending human verification
- Transfer uncertain critical addresses to a manual exception review
Explanation
Correct Answer
C:High-impact numeric ambiguity deserves explicit uncertainty. Forcing a confident amount can create financial or compliance errors even if output JSON is valid.
Incorrect Answers
A:Although semantic checks compare extracted values to business invariants, the choice addresses checking arithmetic consistency in extracted records and not the requested calibrating abstention for unreadable amounts.
B:Feedback-based repair gives the model a specific correctable defect; the deciding condition is calibrating abstention for unreadable amounts, whereas this action concerns targeting a retry at a correctable parse defect.
D:For shipping or compliance, uncertain addresses have operational consequences; this addresses routing low-confidence addresses to human review, while the question is about calibrating abstention for unreadable amounts.
Question 203
A security rule needs to cover Terraform files in any nested infrastructure module. What selector design helps?
- Restrict the rule's paths glob to target only migration files
- Relocate writing-style guidance into a documentation-path rule
- Clarify overlapping conditions and specify one unambiguous policy
- Use a recursive path pattern matching Terraform files across modules
- Refresh the glob to cover nested directories explicitly
Explanation
Correct Answer
D:Recursive globs represent a class of files regardless of module depth. Hand-listing each current folder would be brittle as the infrastructure tree grows.
Incorrect Answers
A:When considering narrowing a conditional rule's file scope, a narrower glob controls the rule's applicability; the requirement for covering nested infrastructure files consistently remains unmet.
B:When addressing limiting editorial guidance to documentation edits, documentation conventions are meaningful when documents are the target; the present task instead calls for covering nested infrastructure files consistently.
C:Conflicting active instructions can make outcomes nondeterministic; preventing contradictory conditional instructions and covering nested infrastructure files consistently require different controls in this situation.
E:Path patterns control rule discovery; in this scenario, matching nested paths in a rule glob is secondary to covering nested infrastructure files consistently.
Question 204
An operations team wants to diagnose why a session reached a certain refund decision. What metadata should persist?
- Store a session identifier and resume its authorized conversation state
- Retain decision timestamps, tool evidence and escalation events with the session
- Compare the persisted operation ID against the payment system before retrying
- Persist message order and essential tool-result linkage along with session metadata
Explanation
Correct Answer
B:Auditable session state links outcomes to observations and approvals. A last-message-only record cannot reconstruct the path that produced a financial action.
Incorrect Answers
A:Session continuity reconnects the agent to relevant history and progress; the action supports resuming a saved agent session but falls short of preserving a session-level decision audit trail in this case.
C:A durable transcript does not guarantee whether the external effect committed; that addresses reconciling external actions after session interruption, whereas this case calls for preserving a session-level decision audit trail.
D:Resumable state includes structured execution history, not merely the final natural-language summary; this concerns checkpointing structured agent progress, leaving the requirement for preserving a session-level decision audit trail outstanding.
Question 205
A typed output field might contain missing information. What schema choice should represent legitimate absence?
- Formalize the tool input schema with a numeric amount field and strict validation where supported
- Tag versions of the result contract and test consumers before deploying schema changes
- Define a nullable field explicitly and use a documented missing-value convention
- Cut down the schema to supported constructs and validate remaining rules in application code
Explanation
Correct Answer
C:A deliberate null representation lets consumers distinguish unknown data from malformed output. Asking the model to invent a value to satisfy a required non-null field harms correctness.
Incorrect Answers
A:Typed schema constraints differentiate numeric amounts from text; its benefits for enforcing typed arguments for a tool call do not resolve the need for representing legitimate missing data in a schema.
B:Structured outputs create a stable interface that downstream systems depend upon; supporting managing downstream compatibility of structured outputs cannot replace the required representing legitimate missing data in a schema.
D:Constrained decoding has a documented supported JSON Schema subset; that serves respecting structured-output schema limitations, whereas the case requires representing legitimate missing data in a schema.
Question 206
A support agent's Messages API response contains stop_reason tool_use and a lookup_order tool block. What must the application do before expecting a customer-facing answer?
- Execute the requested client tool and return a matching tool_result
- Keep the assistant tool_use content block as part of the message history
- Associate and execute both tool calls, returning their respective results
- Provide the result back into the model loop so Claude can decide the next action
Explanation
Correct Answer
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.
Incorrect Answers
B:The next request needs the structured assistant invocation and the matching tool result in a coherent transcript; that may serve preserving the structured conversation transcript, but it is no substitute for completing the client-tool round trip.
C: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 may support handling parallel client-tool calls but cannot alone meet the requirement for completing the client-tool round trip.
D:The application does not simply display raw tool output as the final assistant response; the stated problem requires completing the client-tool round trip rather than resuming model reasoning after a result.
Question 207
A document writer keeps revising style after it has been approved, introducing new factual errors. What workflow boundary helps?
- Score findings by impact and evidence strength before producing recommendations
- Freeze validated facts and limit later passes to approved presentation edits
- Focus on grounding each review in primary evidence rather than inherited generated assertions
- Maintain versions of prompts and rubrics and rerun a fixed regression evaluation suite
Explanation
Correct Answer
B:Separating factual content from stylistic changes reduces accidental meaning drift. Any substantive change should trigger a fresh evidence review.
Incorrect Answers
A:Review output should prioritize well-supported important issues; the option would serve prioritizing findings after independent review, not preserving verified facts during editing as specified in the stem.
C:Recycling earlier model text can propagate one hallucination across every pass; that relates to preventing propagation of unsupported intermediate claims, not the necessary preserving verified facts during editing.
D:Although comparable benchmarks reveal behavior changes when the system evolves, the choice addresses making multipass quality evaluations reproducible and not the requested preserving verified facts during editing.
Question 208
A support integration has separate tools for viewing and canceling an order but their descriptions sound identical. Which revision is necessary?
- Specify explicit amount units and a required currency input
- Present a narrow parameterized order-status operation instead of general SQL
- Specify the supported reason codes as a constrained enumeration
- Distinguish the read-only lookup from the authorized state-changing operation
Explanation
Correct Answer
D:Tool metadata should expose whether a call merely inspects or mutates an order. This prevents the agent from choosing cancellation when only an information request was intended.
Incorrect Answers
A:Financial tools need unambiguous units and typed fields; disambiguating monetary units in tool calls and clarifying read versus write tool intent require different controls in this situation.
B:A purpose-limited interface minimizes injection and unauthorized data access; the unsatisfied requirement is clarifying read versus write tool intent, not constraining a tool to least-privilege business operations.
C:Enumerated fields constrain inputs to a documented business vocabulary and support deterministic validation; in this scenario, limiting parameters to valid business values is secondary to clarifying read versus write tool intent.
Question 209
A security worker needs to examine configuration but must not run shell commands. How should the parent invoke it?
- Use a specialized subagent with only the necessary non-executing inspection tools
- Demand evidence of which documents and versions the worker actually consulted
- Provide an explicit output contract and completion criteria in the delegated prompt
- Forward condensed relevant findings and source handles rather than full unrelated transcripts
Explanation
Correct Answer
A:Tool access should match the worker's responsibility. Giving Bash by default enlarges the attack and mutation surface for a task satisfied by read-only inspection.
Incorrect Answers
B:Source and version metadata distinguish inspected facts from memory-based guesses; this concerns validating the evidentiary basis of a subagent answer, leaving the requirement for preventing command execution in a read-only worker outstanding.
C:A subagent needs a bounded objective, format and acceptance test; making subagent completion testable and preventing command execution in a read-only worker are distinct responsibilities under these constraints.
D:Selecting task-relevant evidence controls token growth without erasing provenance; the action supports controlling delegated context size but falls short of preventing command execution in a read-only worker in this case.
Question 210
A support session's policy version changed during a long interaction. How should a compact summary handle this?
- Carry forward approval identity, timestamp and supporting policy evidence
- Carry forward a concise referenced policy finding and its source version
- Record policy version and recheck rules that govern pending decisions
- Keep pending approvals, blockers and next required actions explicitly
Explanation
Correct Answer
C:Summaries are snapshots rather than proofs of current policy. Version-aware revalidation avoids applying a superseded refund limit to a later action.
Incorrect Answers
A:High-impact decisions require an audit trail, not just a terse conclusion; this contributes to preserving authorization evidence through summarization without meeting the need for detecting stale policy within condensed state.
B:Cached validated findings reduce repetitive reads while retaining provenance; that serves avoiding repeated retrieval of unchanged evidence, whereas the case requires detecting stale policy within condensed state.
D:Pending workflow state is essential for safe resumption; its benefits for keeping unresolved control gates in condensed context do not resolve the need for detecting stale policy within condensed state.