Why the jar fills slowly
A model can look good by luck on a small sample. Vapnik–Chervonenkis theory puts a number on that luck: with n tokens and a model of capacity d, the true score can sit below the measured one by at most ε, with 95% confidence.
ε = √d(ln 2n/d + 1) + ln 4/δn
Émile fills the jar with AUC − ε, never with the raw score. Good model, thin sample: jar stays empty. That is the point.