Silent ML Defects

A thirty-day rolling label turns 3,275 rows into about 110 real examples

Every row stays causally valid and the temporal split stays correct, which is why no leakage check flags this, while the independent observations behind each confidence interval fall by a factor equal to the window length.