Question 260
A customer claims an urgent exception to a standard no-refund policy. Which safe action preserves service quality?
- Disambiguate the intended action and confirm authorization before choosing either operation
- Transfer to the security or fraud workflow and suspend sensitive actions
- Capture the exception rationale and route it through the authorized review process
- Seek authorized identity clarification before account-specific action
Explanation
Correct Answer
C:Unusual circumstances may justify discretion but need a human or policy-sanctioned decision. Automatic denial or approval without the documented exception path may be inappropriate.
Incorrect Answers
A:Destructive and reversible operations are not interchangeable; this contributes to clarifying an ambiguous destructive request without meeting the need for processing policy exceptions through approval.
B:Fraud indicators change the permitted handling of customer accounts; that serves routing suspected fraud out of standard support, whereas the case requires processing policy exceptions through approval.
D:Missing identity is a safety-critical ambiguity; its benefits for escalating unresolved customer identity do not resolve the need for processing policy exceptions through approval.
Question 261
A report generator should abandon publication when mandatory supporting evidence is unavailable. Which workflow policy fits?
- Use a fail-closed publication decision on missing mandatory evidence
- Append a concise case summary, relevant evidence and the escalation reason
- Encode external-input and tool-pending states separately
- Represent compensation or recovery actions for partial multi-step failure
Explanation
Correct Answer
A:If a required validation fails, the pipeline should report incomplete work rather than produce a polished but unsupported artifact. Explicit fail-closed behavior prevents accidental release.
Incorrect Answers
B:A meaningful handoff preserves the context the reviewer needs to act, avoiding repeated questioning and incorrect policy decisions; the stated problem requires failing closed on an unmet evidence requirement rather than transferring sufficient escalation context.
C:Different waits have different timeouts, retries and authorization requirements; a method suited to distinguishing human and tool handoff states does not meet the need for failing closed on an unmet evidence requirement.
D:For recovering from a partially committed workflow, external systems rarely offer one transaction across all tools; however, this requirement concerns failing closed on an unmet evidence requirement.
Question 262
A human approver receives only the proposed refund amount with no customer evidence. What must the review interface include?
- Insist on an authenticated human approval before the mutation is dispatched
- Show policy basis, supporting records, and the requested action before approval
- Sort outcomes by material impact, uncertainty and business risk
- Tie approval to immutable action parameters and require reapproval after material changes
Explanation
Correct Answer
B:Meaningful approval requires adequate context to judge eligibility and consequences. A bare amount invites rubber-stamping or incorrect authorization.
Incorrect Answers
A:Approval must be checked at the controller boundary rather than inferred from model confidence; that relates to enforcing human authorization for high-value actions, not the necessary giving human reviewers sufficient decision evidence.
C:Triage uses the cost of errors to allocate scarce review capacity; the deciding condition is giving human reviewers sufficient decision evidence, whereas this action concerns directing human attention to the highest-risk cases.
D:Approval of one transaction should not authorize a modified one; the option would serve tying human approval to exact mutation details, not giving human reviewers sufficient decision evidence as specified in the stem.
Question 263
A repository contains confidential files, and a policy must prevent tool calls from reading them. Which approach is most suitable?
- Capture the tool completion event and record its returned outcome
- Append the validation to the agent-turn or stop lifecycle event
- Emit a blocking control result instead of an informational log event
- Enforce protected-path checks in a pre-tool permission handler
Explanation
Correct Answer
D:A programmatic check can compare the requested path to restricted locations before access occurs. Textual instructions alone provide weaker protection against accidental reads.
Incorrect Answers
A:Post-execution telemetry can contain the result and status of a tool operation; in this scenario, auditing tool execution outcomes is secondary to denying sensitive file reads before access.
B:When considering validating at the end of agent activity, end-of-turn hooks can trigger aggregate checks that would be wasteful after every individual edit; the requirement for denying sensitive file reads before access remains unmet.
C:When addressing distinguishing logging from access denial, emitting telemetry alone does not enforce permission decisions; the present task instead calls for denying sensitive file reads before access.
Question 264
A report synthesizer combines three source summaries but loses which original passages support each fact. What is the right context representation?
- Track policy version and recheck rules that govern pending decisions
- Maintain pending approvals, blockers and next required actions explicitly
- Compress earlier turns into a verified summary that keeps decisions and open actions
- Store claim-to-source mappings alongside the condensed findings
Explanation
Correct Answer
D:Compression should preserve provenance links for verifiable conclusions. A citation-free summary makes later fact checking and correction much harder.
Incorrect Answers
A:Summaries are snapshots rather than proofs of current policy; detecting stale policy within condensed state and maintaining citations during context compression are distinct responsibilities under these constraints.
B:Pending workflow state is essential for safe resumption; the action supports keeping unresolved control gates in condensed context but falls short of maintaining citations during context compression in this case.
C:Context compaction reduces token load while retaining essential state; this can serve retaining critical state while compacting history, but leaves the requirement for maintaining citations during context compression unmet.
Question 265
A tool rejects an invalid account ID format and the agent retries the same request five times. What metadata would improve behavior?
- Return a non-retryable validation error with the offending field identified
- Group the business constraint as nontransient and require a different next step
- Discriminate between a legitimate not-found result from a tool execution error
- Hide secrets in public tool errors and retain details in protected logs
Explanation
Correct Answer
A:Input-shape errors persist until arguments change. Explicit field-level diagnostics help the agent correct the request instead of wasting retries on an unchanged invalid value.
Incorrect Answers
B:Funds or policy failures generally need user or system-state changes, not immediate identical retries; its benefits for avoiding futile retries on business-rule rejection do not resolve the need for distinguishing validation faults from transient errors.
C:A successful search with zero matches is not necessarily an infrastructure failure; supporting separating no-match outcomes from failures cannot replace the required distinguishing validation faults from transient errors.
D:Agent-visible diagnostics should be actionable without exposing credentials; this contributes to sanitizing error information returned to models without meeting the need for distinguishing validation faults from transient errors.
Question 266
A contract extraction process repeats an unsupported claim across reviewers and sometimes misreads financial totals. Which TWO review steps improve independent evidence checking? Choose TWO.
- Have a second pass recompute and confirm contract financial amounts against the original file
- Make independent reviewers inspect source passages rather than inherit earlier generated claims
- Freeze the first extracted financial total and prohibit later passes from questioning it
- Treat matching wording across reviewers as independent evidence without reopening the source contract
Explanation
Correct Answers
A:This choice addresses trading review latency for financial precision: a dedicated verification step can compare amounts and field provenance before acceptance; extra latency can be justified when financial errors are costly.
B:Recycling earlier model text can propagate one hallucination across every pass. Reviewers need direct access to source material to supply meaningful independent checks.
Incorrect Answers
C:Locking the first pass's numbers perpetuates extraction errors. Independent validation against the original document is required before any values can be accepted.
D:Agreement can reflect shared model bias or copied assertions. The reviewers need access to the primary document and must verify financial fields independently.
Question 267
A parent delegates report drafting but keeps asking the worker to decide whether data deletion is authorized. What boundary should change?
- Produce stable source identifiers alongside evidence extracts and claims
- Assign the task with the precise file paths or retrieved evidence required for the task
- Keep approval for destructive decisions with the coordinator or an authorized human
- Insist on structured task results with IDs, status and completion evidence
Explanation
Correct Answer
C:Delegating writing does not imply delegating business authority. The parent should retain policy-sensitive commitments and use workers for bounded preparation.
Incorrect Answers
A:Citation-safe merge requires referential metadata; the deciding condition is retaining authority for irreversible actions, whereas this action concerns preserving source pointers through delegation.
B:Subagents have their own context and should be given task-critical inputs explicitly; the option would serve passing task-specific context into a subagent, not retaining authority for irreversible actions as specified in the stem.
D:Stable identifiers and machine-readable outcomes let the parent distinguish completed assignments from new work; that relates to normalizing delegated task results, not the necessary retaining authority for irreversible actions.
Question 268
A customer-facing agent returns stop_reason refusal to a disallowed request. What should the workflow do?
- Route the refusal through the approved response and escalation policy
- Validate the side effect with an authoritative follow-up read before promising completion
- Correlate and execute both tool calls, returning their respective results
- Interpret the output as truncated and continue or retry with an adequate token budget
Explanation
Correct Answer
A:A refusal is a distinct termination state. Treating it as a tool request or silently retrying the prohibited content would ignore safety and workflow policy boundaries.
Incorrect Answers
B:A tool's acknowledgement can be stale or incomplete relative to durable state; verifying externally visible side effects and handling refusal as a distinct state require different controls in this situation.
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; the unsatisfied requirement is handling refusal as a distinct state, not handling parallel client-tool calls.
D:A max_tokens stop means the generation reached the response budget, not that the report passed validation; in this scenario, detecting token-limit truncation is secondary to handling refusal as a distinct state.
Question 269
A new customer appears with an unrelated support issue. Should the old customer's session be reused?
- Define a new fork from the last accepted checkpoint for revised analysis
- Create an alternative branch of from the same validated baseline and keep branch histories separate
- Start a separate session with its own customer and authorization context
- Match the fork against the current codebase and rerun relevant checks
Explanation
Correct Answer
C:Separate sessions prevent accidental disclosure and policy contamination between customers. Carrying unrelated prior history into a new case creates privacy and correctness risks.
Incorrect Answers
A:Branching preserves what was previously considered while enabling a corrected trajectory; this can serve revising an analysis without erasing history, but leaves the requirement for isolating unrelated customer conversations unmet.
B:A common starting context makes alternatives comparable; that addresses comparing alternatives from a common session baseline, whereas this case calls for isolating unrelated customer conversations.
D:A successful result on an older branch does not prove current compatibility; validating branch changes before merge and isolating unrelated customer conversations are distinct responsibilities under these constraints.
Question 270
A team copies hundreds of examples into every call and exceeds latency targets. What optimization fits?
- Preserve example-set versions and compare outcomes on a fixed test suite
- Build in an example that correctly returns an unknown or escalation category
- Provide contrasting high-severity and low-severity labeled cases with rationales
- Retain a small representative few-shot set and evaluate the trade-off empirically
Explanation
Correct Answer
D:Large demonstrations consume context and tokens. Selective examples can preserve important decision boundaries while reducing request size.
Incorrect Answers
A:Versioned few-shot experiments reveal the impact of changes; supporting measuring few-shot changes with controlled evaluations cannot replace the required controlling context cost from examples.
B:Boundary examples teach that abstention is acceptable; this contributes to calibrating abstention with a boundary example without meeting the need for controlling context cost from examples.
C:Contrastive examples teach the decision boundary between categories; that serves distinguishing severity classes with contrastive examples, whereas the case requires controlling context cost from examples.