In “How surprising?”, an outcome with chance p took log2(1/p) yes/no questions to pin down, so likely outcomes are found fast. You set up your questions for the chances planned_chances, but outcomes really happen with true_chances. Write wasted(true_chances, planned_chances): the extra questions per outcome, on average, compared with having planned for the truth.
| A | B |
|---|---|
| 0.5 | 0.5 |
| A | B |
|---|---|
| 0.99 | 0.01 |