I've been listening to talk radio here in the UK, and it's a constant stream of disappointed kids/teachers/parents phoning in. Some of the stories are horrendou

I've been listening to talk radio here in the UK, and it's a constant stream of disappointed kids/teachers/parents phoning in. Some of the stories are horrendous, people expecting AAA getting CCD and similar. It seems unlikely that you'd ever have three teachers predicting you to get top marks, but you end up with that on results day. But about the algorithm, it seems to be an impossible task. You cannot satisfy the requirement of the gross statistics being roughly the same, ie little grade inflation on average, while also identifying talented individuals. The outline of the algorithm seems to be that you basically take the gross stats for earlier years from a school, and then you take the teacher's guesses about who will do well, and make up a distribution of grades around that. This is probably what most people would do if forced into such an exercise. And you can add wrinkles to it like a bit of grade inflation, adjustment for overconfident predictions, and so on. But you can't get that piece of information that you really need, which is a mix of signal and noise about the true distribution of the kid's abilities. If you have a talented year in a low performing school, you'll never know. And vice versa. The year before I finished the IB, two kids got the top 45/45. That's normally achieved by something like 1/400 kids worldwide, and it was a class of maybe 30 kids. That would never have been awarded without examinations. In addition the other presumed goal, of having the grades as close to the predictions as possible, give the algo designer an incentive well known from machine learning: guess safely. Squish the distribution on the edges, that way you're not off much. This is bad for people who shoud be getting As, because many will get Bs. And it's good on the other end, though I don't know where the lump ends (C or D?).