After years of watching smart teams mistake sampling for safety, I no longer ask how many AI tests we ran, only which failures we have made impossible by design.