Answer out loud before looking back.
One. Why is confident invention a property of the mechanism rather than a bug?
Two. A model is right ninety-nine times in a hundred. What do you need to know before you can say whether that is acceptable?
Three. Why might a model score well on a public benchmark for reasons that do not help you?
Four. What is wrong with using a model's agreement as evidence that your design is sound?
Five. An agent reports that all tests passed. What do you ask for, and why is the summary not it?