A classifier learns from targets: for a photo of a dog, the usual target is 100% dog and 0% for everything else (“Colors into numbers”). But its chances come from softmax (“Scores into chances”), which only reaches exactly 100% when the dog score is infinitely far above the rest, so a model chasing that target pushes its scores apart forever and grows overconfident (“Score a forecaster”). Write target(correct, classes, doubt): the chances to train toward for a photo whose right answer is class number correct (0 is the first) out of classes, keeping doubt (0 to 1) of your belief spread around.