Draw inside the distribution.
Clean templates produce sharp activations because their geometry resembles the examples the model was designed around.
THE OOD PROBLEM
Confident answers.
Uncertain foundations.
Draw something the model knows—or something it has never seen. Watch the same softmax machinery turn both into a neat answer.
INFERENCE 02
READYInside the approximation
INPUT
35 features
35 features
OUTPUT
10 classes
10 classes
Step outside it.
Make a shape that is not a digit. The classifier still has to choose one of ten labels.
Read confidence critically.
A high softmax score is a preference among known labels—not proof that the input belongs to the world those labels describe.