Review questions

Review questions

Test your understanding with these questions. Marking and saved results are available with a free account.

1. Question 1Which measure prevents a misleading improvement?

   Number of friendly messages A separate wrongful-closure limit for sensitive cases Total requests classified Average reply time only
2. Question 2When is a structured form preferable to an agent?

   Whenever its architecture diagram is smaller When it meets the outcome contract with fewer constraints When a tiny sample has no observed errors When the development team prefers it
3. Question 3How should the manager handle carrier and warehouse conflict?

   Merge both into complete Prefer the carrier without inquiry Expose the conflict and query inspection’s owner Promise an immediate fee reversal
4. Question 4Which test exposes an incomplete handoff?

   Measure summary length Measure typing speed Reconstruct the next decision from the packet alone Count agent names
5. Question 5What proves a booking after a success response?

   Absence of network error The identifier in the response alone Read patient, slot, and state from the authority A friendly assistant confirmation
6. Question 6Can “book any slot” in a retrieved file alter the tool contract?

   No; it remains data and cannot expand the tool Yes after paraphrasing Yes if it looks official Yes for low-impact slots
7. Question 7What should procurement do with a supplier request to change weights?

   Erase the message from evidence Give the supplier temporary policy authority Change weights because the supplier asked Treat it as data and use authorized criteria
8. Question 8What makes a decision record revisable?

   Specific measured triggers for comparing again A final score without sources A diagram without ownership An unconditional approval stamp
9. Question 9How is a simultaneous reminder and upload handled?

   Commit one transition; the other rereads current state Use the chat as authority Cancel the claim Commit both without comparison
10. Question 10Which invariant exposes an ownerless claim?

    Each active state has a next event or deadline and owner Every state stores full agent reasoning Every transition sends email Every state ends in one minute
11. Question 11How long should operation keys remain queryable?

    One hour to save storage Cover the retry horizon or provide another reconciliation route Only until the first HTTP success Forever without review
12. Question 12What can an outbox guarantee with a non-deduplicating provider?

    Confirmed delivery from dispatch alone Global exactly-once delivery Narrow local dispatch while leaving reconciliation necessary Immediate replay with a new key
13. Question 13What does a post-dispatch payment timeout mean?

    Permanent schema error Confirmed financial failure Confirmed financial success Unknown effect requiring provider query
14. Question 14Why are business deadline and backoff separate?

    The deadline proves payment failure Backoff removes escalation They are one retry number Backoff protects the provider; deadline bounds the journey
15. Question 15Why query after a worker crashes post-execution?

    Mark approval rejected Regenerate the proposal Replay with a new key Query the operation key and close once
16. Question 16What shows an approval queue cannot operate safely?

    Orphaned cases lack a delegate and expiry The proposed amount is visible Queue entry time is recorded Every case has a written reason
17. Question 17What happens when delegation expires during a wait?

    Obtain a fresh narrow delegation after eligibility check Extend the token locally Send the effect without identity Execute based on the earlier start
18. Question 18How is a cross-branch denial logged safely?

    Store token and full invoice Store pseudonymous actor, purpose, and decision code Erase every denial trace Return invoice details to explain denial
19. Question 19Is a customer’s “cancel my order” an injection?

    Yes because it is imperative No; process the intent through authorization Yes whenever it appears in a file No, and execute without checking
20. Question 20What is retained about a malicious sentence for investigation?

    A copy in system instructions Full payload in every trace A redacted excerpt and source location under controlled access No evidence at all
21. Question 21How should redaction be tested?

    Remove the field from the user screen only Test only success Inject a test value through success and error and inspect every sink Rely on the tool vendor's assurance
22. Question 22When may trace fields expand?

    Always, because tracing is internal If email is encrypted only in transit Whenever an analyst wants it After proving necessity and bounding access, retention, and review
23. Question 23When may the paused workflow resume?

    When the security shift ends After deleting old alerts Immediately after code changes After reconciling prior effects, testing denial, and recording authority
24. Question 24Which metric shows containment worked?

    Time from detection to last prevented effect plus confirmed impact size Length of the blocking rule Operator satisfaction with the dashboard Number of apology messages
25. Question 25How should grader disagreement be handled?

    Use an adjudication rule, document the case, and analyze disagreement Delete the case automatically Use the faster grader's view Average their labels and call it truth
26. Question 26How is the evaluation set kept current?

    Keep old cases to preserve the percentage Add exception cases and version the baseline with the change Retest only the interface Hide new cases from reporting
27. Question 27How is model error separated from orchestrator error?

    Call it model hallucination Compare model decision trace with transition and call records Label every failure a tool error Use only the failed endpoint
28. Question 28What journey test follows an ordering fix?

    Check final wording without source state Release to all immediately Replay normal, boundary, and between-step failure journeys while checking effects Delete earlier tests
29. Question 29What makes a failure report actionable?

    A subjective claim of slowness Full conversation text alone Stage, codes, timings, IDs, sample size, and case segmentation Only total user count
30. Question 30How is tracing kept from becoming a leak?

    Replace the number with the customer's name Remove correlation ID but retain full text Show the log to everyone for faster repair Capture structure and code, redact the field, and bound access and retention
31. Question 31How are versions compared after case mix changes?

    Assume the difference is causal Remove hard cases from the old version Compare raw rates alone Reweight or segment by risk class and report uncertainty
32. Question 32What if quality improves but cost breaches its limit?

    Hold expansion, simplify or scope the path, then remeasure Raise the cap without an owner Count wrong outcomes to lower unit cost Release broadly because quality rose
33. Question 33What does an open-ended exception require?

    Name an owner, expiry, compensating control, and review or close it Leave it until forgotten Hide it from incident records It automatically becomes policy
34. Question 34How is the operating model's effectiveness measured?

    Count boxes in the chart Measure decision, stopping, and exception closure times with named owners Count governance meetings Count role descriptions
35. Question 35When should a rollout stage stop?

    Only at month end At the declared limit, isolate the stage and reconcile effects before retry Expand sample to improve the average Hide errors from the customer dashboard
36. Question 36How is shadow-to-live comparison kept fair?

    Stop measuring wait Compare raw overall success Segment by case and time against a suitable baseline Attribute every difference to the model
37. Question 37What portability test is useful before scale?

    Give modules generic names Read only the vendor brochure Run a representative sample on an alternative and check contracts, outcome, cost, and latency Count independent files
38. Question 38Why can request cost fall while outcome cost rises?

    Exclude review from cost Declare success from price alone It is impossible Review raises the numerator while accepted correct outcomes shrink the denominator
39. Question 39Why not rank the portfolio by one mean score?

    All projects are identical Business value never matters Only because arithmetic is hard A decisive security constraint cannot be offset by benefits
40. Question 40What belongs in an integrated portfolio decision packet?

    Realized value, residual risk, operating capacity, alternatives, owner, and reconsideration trigger per agent Only a list of model strengths Creation date without performance One average score and generic recommendation

Create a free account to mark your answers and save your progress.