AI benchmark leaders look unbeatable. In real workflows, smaller specialized models can still win. The reason exposes what public leaderboards fail to measure. #AI