Question 40
A policy team wants every support answer mentioning regulated advice to show an escalation notice, even if the model forgets. Which solution provides a stronger guarantee?
- Enforce the condition in application logic before releasing the response
- Train customer agents to check only a random monthly sample
- Add the notice as an optional style tip in the prompt
- Increase the model's creative sampling to make notices more prominent
Explanation
Correct Answer
A:Deterministic checks can block or augment responses that match the defined escalation criteria. Instructions alone cannot ensure every generated answer complies.
Incorrect Answers
B:Sampling is useful for audits, not per-answer enforcement. The requirement applies to every qualifying response.
C:A language-model instruction can be overlooked. Mandatory behavior needs a coded guard or approval gate.
D:Greater output variability does not improve guaranteed compliance. It can make consistency worse.
Question 41
A support agent has completed all checks but a customer asks for a policy exception outside its authority. What is the most appropriate terminal outcome?
- Escalate with evidence and clearly state that approval is pending
- Approve the exception because the customer is unhappy
- Silently discard the conversation and return no status
- Send the case to a search worker with full refund privileges
- Use a generic search result as if it were a signed management approval
Explanation
Correct Answer
A:Handoff preserves the case history and avoids representing an unauthorized exception as settled. A human or authorized system decides the exception.
Incorrect Answers
B:Customer sentiment does not grant transactional authority. The rule still requires an approved exception process.
C:Dropping the case destroys context and leaves the customer without a path to resolution.
D:A retrieval specialist does not become an authorizer by receiving the issue. Privileged approval belongs to a designated reviewer.
E:Retrieval provides information, not authority to waive policy. The approval must come from the designated role and be recorded.
Question 42
A service organization requires a record of why a human approved an unusual refund. Where should this evidence be captured?
- An auditable decision record linked to the tool action
- Only in a friendly closing message to the customer
- Only in the model's invisible reasoning text
- By asking the model to remember the case indefinitely
Explanation
Correct Answer
A:The record should retain the initiating request, policy trigger, approver decision and executed result. This supports investigation after the conversation ends.
Incorrect Answers
B:A closing message does not capture the complete control evidence and may omit staff identity or policy exception details.
C:Internal model deliberation is not a reliable approval ledger. Audit trails require explicit application-owned records.
D:Model memory is not a durable approval system. Decisions need stored, controlled evidence.
Question 43
A returns workflow sometimes publishes an answer even when a downstream refund API times out. What status behavior should replace this?
- Assume every timeout means the refund succeeded
- Represent the action as unconfirmed and check or reconcile backend state
- Assume every timeout means the refund failed and immediately retry
- Remove the timeout handling so the agent can continue naturally
Explanation
Correct Answer
B:A timeout does not establish whether a write succeeded. The workflow should prevent false success statements and resolve ambiguity before retrying.
Incorrect Answers
A:A network timeout contains no proof of completion. Claiming success can mislead customers and audit systems.
C:An earlier request may have succeeded despite the lost response. Blind retries can duplicate money movement.
D:Ignoring uncertainty masks a transactional failure mode. Explicit recovery or human follow-up is needed.
Question 44
A legal drafting assistant may create recommendations but only counsel may approve a filing. Which workflow boundary is most appropriate?
- Ask a second model to impersonate counsel for the final sign-off
- Skip the review whenever the document is shorter than a page
- Permit the drafting agent to publish if its own confidence is high
- Require human release approval after automated drafting
Explanation
Correct Answer
D:The system separates preparation from authorized publication. Counsel reviews the actual artifact and owns the final legal action.
Incorrect Answers
A:A second generation pass can review quality but cannot replace a designated accountable human.
B:Document length is not the trigger for authorization. Even short filings can have significant legal effect.
C:Self-reported certainty does not confer legal signing authority. A human approval gate is a role requirement.
Question 45
A team needs a guaranteed approval checkpoint before a write tool, but the assistant sometimes skips a textual reminder. Which design principle applies?
- Put compulsory checkpoints in execution controls rather than only prompts
- Repeat the reminder ten times and assume the guarantee is now absolute
- Hide the write tool from logs after the action completes
- Use a more expensive model instead of an approval service
Explanation
Correct Answer
A:The executor can hold or deny a call until an external approval record exists. Generated intentions must not bypass deterministic authorization.
Incorrect Answers
B:Repetition may increase compliance probability but does not turn probabilistic behavior into enforcement.
C:Suppressing telemetry weakens accountability and does not protect the pre-execution boundary.
D:Model capability does not substitute for access-control or release-policy enforcement. The approval decision remains external.
Question 46
A support bot receives a harassment complaint that its refund policy does not cover, and staff must review the evidence. What should the bot return?
- An unrelated generic FAQ with no escalation record
- A refund transaction merely to end the interaction
- A decisive accusation generated from the complaint alone
- A transparent handoff state with a concise, factual case summary
Explanation
Correct Answer
D:A structured escalation preserves the customer's report for human review. The assistant should not manufacture a policy determination outside its scope.
Incorrect Answers
A:Deflection loses the evidence and fails the stated need for staff review. The case must enter the designated process.
B:Financial compensation is unrelated to adjudicating the complaint. It also creates an unapproved side effect.
C:A claim from one party is evidence to review, not a verified finding. Premature judgment risks unfair or unsupported outcomes.
Question 47
An agent can choose a response path but must not change the required identity-verification stage. What division of labor is appropriate?
- Enforce identity checks before permitted downstream steps
- Replace the fixed gate with a high-confidence classification label
- Make all actions manual even when they are read-only and safe
- Let the model omit verification when the conversation sounds trustworthy
Explanation
Correct Answer
A:The architecture combines deterministic control where policy is absolute with flexible reasoning for authorized task routing.
Incorrect Answers
B:A probabilistic score cannot enforce an authorization prerequisite. The gate needs an explicit trusted decision.
C:Requiring staff for every step is unnecessary when the only fixed requirement is a verification gate before sensitive work.
D:Linguistic plausibility is not proof of identity. Skipping a hard prerequisite violates the workflow contract.
Question 48
A Claude Code workflow must prevent execution of a forbidden shell command even if prompted to ignore policy. Which enforcement point fits?
- Use a PostToolUse hook to complain after the command has run
- Use a PreToolUse hook to inspect and block the proposed tool invocation
- Rely on a final answer review after the session finishes
- Use only a CLAUDE.md sentence stating the command is prohibited
Explanation
Correct Answer
B:A pre-execution hook can evaluate tool names and inputs at the boundary. It creates a policy control separate from model-generated instructions.
Incorrect Answers
A:A post-event hook is too late to prevent side effects. It is appropriate for observing or reacting to completed operations.
C:Text review cannot undo a disallowed command that already executed. The action needs a preflight gate.
D:Project instructions influence tool selection but cannot guarantee refusal in every run. Enforcement must occur before execution.
Question 49
A tool returns filenames with inconsistent path separators, confusing a downstream parser. Where can you normalize the observation?
- Change the user's initial message each time a file is returned
- Transform the result through a PostToolUse hook
- Use PreToolUse to alter the already-produced result
- Depend on the model to mentally repair malformed data every time
Explanation
Correct Answer
B:A post-use hook operates on results after execution. It can normalize returned metadata while preserving the actual outcome and status.
Incorrect Answers
A:Editing user intent is an inappropriate repair for executor output. Normalize at the observation boundary.
C:A pre-hook runs before tool execution and cannot inspect output that does not yet exist. That hook is for proposed inputs and permissions.
D:Probabilistic correction can fail and adds noise. A deterministic transformation is more reliable for a known format problem.
Question 50
A compliance team needs an audit entry whenever a tool mutates a tracked file. What is a suitable hook strategy?
- Audit tool, target and outcome with PostToolUse
- Disable logging to avoid exposing tool paths to reviewers
- Make the assistant write an audit poem at the end
- Log only the user's intention before the tool executes
Explanation
Correct Answer
A:The hook observes completed work and records what actually happened. Pairing target and status supports traceability of changes.
Incorrect Answers
B:Avoiding logs defeats the stated compliance need. Sensitive audit records can be protected with access controls.
C:Generated prose is not a structured event log and may omit failed writes. Audit data should be collected from execution events.
D:Intent does not prove a file was changed. An outcome record is needed to distinguish attempted from successful mutations.