Appendix C. Exam day guide

Before you start

  • Prepare a quiet place, a stable internet connection and a charged device.
  • You have 35 minutes for 25 questions: an average of 1 minute 24 seconds each.
  • Check your progress: aim to complete question 10 by the 12-minute mark and question 20 by the 24-minute mark. Mark difficult questions and move on.

How to read a question

  1. Read the question first. Check whether it asks for the most appropriate answer, two suitable answers or a sequence.
  2. Find the bold conditions. These distinguish correct answers. Interpret what they rule out. External publication suggests verification and approval; hard-to-reverse actions suggest human approval; recent information suggests current evidence; high volume and fast responses suggest small models or batch processing.
  3. Combine at least two conditions. A good question has only one option meeting both. An option meeting just one is a trap.
  4. Judge options by principles. Attach a key term to explain why each action is right or wrong: knowledge cutoff, inherited permissions, baseline and so on.

Signals that help eliminate wrong answers

Warning sign Example Related chapter
Treating specificity or fluency as accuracy "The numbers are specific, so I trust them" Chapter 1
Treating consistency as truth "Repeated questions give the same answer, so it is correct" Chapter 1
Solving problems through emphasis "Write 'never make mistakes' three times" Chapter 2
Unnecessarily heavy tools "Start with an agent or API for a one-off task" Chapter 3
External actions without approval "Delegate sending as well, because that is faster" Chapter 3
Confusing settings dimensions "Disable training and administrators cannot see it" Chapter 4
Skipping verification "Accuracy is high, so remove human review" Chapter 4
Treating usage as outcomes "Question counts increased, so it succeeded" Chapter 5

Tips by question format

  • Single-select: The longest or most cautious option is not necessarily correct. Choose by whether all conditions are met.
  • Multiple-select: Check how many answers are required. Selecting even one too many or too few earns no points.
  • Statement evaluation: Even if the initial judgment is right, mark No when the reason following "because" is wrong. Check judgment and reason separately.
  • Ordering: Ask which earlier result each step uses as input. Steps establishing comparison foundations, such as criteria, baselines and success conditions, generally come first.
  • Matching: Eliminate definite pairs first. If one choice may remain unused, look for the choice that is not a core requirement of any item.

The final five minutes

  • Return to marked questions and reread the bold conditions.
  • Change an initial answer only when you have identified a new condition, not out of vague anxiety.
  • Check that no questions remain unanswered.