Question 260
A multi-day research agent must survive context compaction and process restarts while avoiding repeated external writes. Which TWO state records matter most? Choose TWO.
- An unlimited transcript with no indexing or selection strategy
- Every greeting and stylistic aside in the active context
- A durable ledger of completed work, remaining tasks and source references
- A transaction execution record with identifiers and verified outcomes
- A model-generated guess about which operations succeeded
Explanation
Correct Answers
C:The ledger preserves progress and evidence beyond prompt truncation. It lets the resumed agent choose the next necessary investigation.
D:Conversation history alone may not show whether a write committed before interruption. Durable transaction state prevents duplicate effects.
Incorrect Answers
A:Raw history may overflow context while hiding critical facts. Curated external state is more useful.
B:Conversational filler does not establish progress or transaction outcomes. It is low-priority state for technical recovery.
E:Speculation is not a trustworthy execution ledger. Recovery must consult actual backend evidence.
Question 261
A support case has changed account status since the conversation began. How should the agent treat the old status retained in its summary?
- Delete the entire case history instead of refreshing the mutable fact
- Treat the model's last answer as stronger evidence than the database
- Assume all summarized facts remain true indefinitely
- Recheck mutable account state at the point of decision
Explanation
Correct Answer
D:Summaries preserve past observations, not guaranteed current reality. Financial or permission decisions should use authoritative live state.
Incorrect Answers
A:Other facts may remain relevant. Targeted refresh avoids unnecessary context loss.
B:Generated prose is not an authoritative account record. The current system of record should prevail.
C:External data can change independently of the session. An old account status may no longer authorize action.
Question 262
An agent repeatedly loses the primary success criterion during long code refactors. What checkpoint artifact would help?
- A decorative summary of how the project feels
- A list of unrelated repository statistics
- A task plan with acceptance tests and affected files
- An unstructured dump of every model thought
Explanation
Correct Answer
C:A progress file preserves goals that might disappear during context compression. It also helps resume after interruptions.
Incorrect Answers
A:General sentiment cannot guide file changes or tests. Concrete criteria are essential.
B:Statistics that do not relate to the task consume space without preserving the critical objective.
D:Verbose internal deliberation does not create a stable acceptance record. The checkpoint should be concise and verifiable.
Question 263
A long-running agent decides to compact context. Which information should receive highest preservation priority?
- Private credentials copied from tool logs
- Requirements, key outcomes, open decisions and source refs
- Repeated greetings and rhetorical filler
- Every speculative hypothesis as an established fact
Explanation
Correct Answer
B:These facts govern future actions and prevent unsafe repetition. Less relevant conversational detail can be compressed more aggressively.
Incorrect Answers
A:Secrets should not be replicated into long-term conversational summaries. The agent should use secure runtime access instead.
C:Polite exchanges do not determine work completion or authorized actions. They are low priority for technical continuity.
D:Compaction must distinguish hypotheses from observations. Flattening uncertainty can mislead later decisions.
Question 264
A customer asks to close the account but has two linked profiles with different balances. What is the safest first response?
- Confirm the account and authorization before closure
- Close the account with the smallest balance as a heuristic
- Close both profiles to avoid missing the requested one
- Assume the oldest profile is intended without asking
Explanation
Correct Answer
A:The target is ambiguous and the action has lasting consequences. Identification and approval must precede mutation.
Incorrect Answers
B:Balance does not establish customer intent. Guessing the target is unsafe for destructive operations.
C:That broadens the action beyond consent and can harm an unrelated account. Ambiguity is not permission.
D:Chronology is not a reliable identifier. The workflow should request the missing distinction.
Question 265
An assistant cannot verify whether a customer is eligible for a high-value exception. The policy requires staff approval. What should it do?
- Use self-assessed confidence as substitute for approval
- Escalate with evidence and an unresolved eligibility state
- Allow the assistant to declare itself the authorized reviewer
- Grant the exception and request verification later
- End silently because the policy is difficult to interpret
Explanation
Correct Answer
B:The assistant should preserve the case and avoid pretending approval has occurred. Staff review is the documented decision point.
Incorrect Answers
A:Confidence is not a role permission. The workflow must follow the defined human checkpoint.
C:A model cannot assign itself a protected role. The exception decision belongs to the designated human process.
D:Authorization after execution defeats the policy gate. A high-value exception needs prior approval.
E:A structured handoff gives the user a resolution path. Silence loses the case and its evidence.
Question 266
An invoice classifier cannot decide whether a partly visible digit is 3 or 8. The field controls a payment. Which outcome fits reliability requirements?
- Choose 8 because it produces a more plausible total
- Remove the field and approve payment anyway
- Mark the field uncertain and request verified review
- Write both digits in one numeric field and call the record valid
Explanation
Correct Answer
C:Material ambiguity must be exposed rather than silently resolved. The payment can proceed only after the value is verified.
Incorrect Answers
A:Plausibility is not proof of the source digit. A financial action should not rely on a guess.
B:The uncertain amount remains critical. Omitting it bypasses the required evidence.
D:An ambiguous string does not meet a numeric contract. The system needs a review state.
Question 267
A support assistant is asked for a legal determination that it is not authorized to make. What response behavior is appropriate?
- Present a definitive legal verdict in a confident tone
- Invent citations to make the verdict look official
- Explain the scope limit and offer a route to qualified human review
- Treat the customer's urgency as permission to bypass review
Explanation
Correct Answer
C:A clear handoff avoids unsupported authority claims. The system can still provide relevant factual context within its permitted role.
Incorrect Answers
A:Fluent language does not supply legal authority or full case analysis. The assistant should not overstate its role.
B:Fabricated sources make the outcome less trustworthy. Uncertainty should be addressed honestly.
D:Urgency can affect prioritization but does not grant professional authority. Escalation remains necessary.
Question 268
A user reports an address that conflicts with an authenticated account record. Shipping labels are about to be printed. What should the agent do?
- Ignore the conflict in the customer communication
- Print labels for both addresses
- Use the chat value immediately because it is newer
- Pause label generation and resolve the address mismatch
Explanation
Correct Answer
D:Conflicting addresses create delivery and privacy risks. The current authoritative record should not be overridden by an unverified chat claim.
Incorrect Answers
A:The user must know the action is pending resolution. Hiding ambiguity can cause missed deliveries.
B:Multiple labels could send sensitive shipments to the wrong place. A single verified destination is needed.
C:Recency alone does not establish authorization or correctness. The source and identity checks matter.
Question 269
A chatbot must decide whether to continue troubleshooting or hand a complex outage to staff. What should govern escalation?
- Defined thresholds for severity, retries and authority
- Whether the assistant's prose sounds sufficiently positive
- A random selection to balance staff workload
- The number of exclamation marks in the customer's message alone
Explanation
Correct Answer
A:Measured criteria support consistent handoffs. A case exceeding safe automation limits should enter an accountable human workflow.
Incorrect Answers
B:Optimistic tone does not indicate outage severity. Operational evidence and policy should drive routing.
C:Random routing ignores customer risk and task complexity. It can delay urgent cases.
D:Sentiment can be informative but does not establish technical impact. It is not a sufficient escalation rule.
Question 270
A customer asks for a charge reversal but supplies no transaction reference and several payments match. How should the workflow respond?
- Reverse the largest charge as a likely match
- Resolve the unique transaction ID before changing funds
- Reverse every charge and let finance clean up later
- Use the model's internal memory to invent an invoice number
Explanation
Correct Answer
B:A specific target is necessary to prevent reversing the wrong payment. The agent can gather additional authorized information.
Incorrect Answers
A:Amount-based guesswork is unsafe when multiple transactions are possible. A financial mutation requires exact identity.
C:Multiple side effects go beyond the request and increase reconciliation costs. Safe targeting should come first.
D:An invented identifier cannot ground a backend action. Verification needs trusted account data.