This is the failure mode that belongs to embodiment and to nothing else. When the arm moves fast, the robot's balance controller adjusts the legs to compensate the reaction. That leg work is real energy, and on this robot some of it lands on the same rail the arm is measured through. So the arm's power reading includes work the arm did not do. Nothing in the electrical signal says so — the number is plausible, the residual is unremarkable, and the model quietly fits the wrong thing. What exposed it was not a better ammeter. It was the inertial sensor: trials where body angular velocity was unusually variable were the contaminated ones, and 104 of 1,017 were rejected on that basis.