A practice trainer for the error decomposition in learning from data. Each round gives a hidden rule Y = f(X) + ε and a noise distribution, freezes X at one value, and asks for the center and spread of the pile of y’s there and its full conditional distribution; or scores a rival guess by its expected squared miss, split into a reducible part and the irreducible noise floor; or shows five fits’ predictions at one input and asks for the pile’s center, bias, variance, and expected test error. Type a fraction (any equivalent form counts, a leading minus is allowed), get a diagnosis or a worked table either way, and generate fresh numbers to try again.

Freeze x, read the pile: its center is the rule, its spread is the noise. Then score a rival guess — and a pile of fits. Fresh numbers each round.