A machine can report what it actually did, in joules.
Software asserts. Energy is spent. A power trace is the one account of a machine's activity that its own code cannot simply write, which makes energy the natural substrate for a machine that has to demonstrate its own behaviour rather than claim it. The question this topic asks is not whether that works, but how far it reaches, and the answer turns out to depend on the body. The same hidden computation is plainly visible on one robot and entirely invisible on another, and no adversary chose which.
Three bodies, one question. On an autonomous mobile robot the motors draw 16.6% of system power while the GPU alone draws 37.3%, compute is the larger term, so hidden computation shows. Drive the same robot manually with navigation and detection off and the motors jump to 64.0%. Switch the meter to the motor rail and the same work vanishes completely, because that plane never carried it. The reading was honest every time; it was answering a different question.
The open half of the topic, and it is open in public. Two motors off one line differ by about 21.5% in the coefficient that could serve as an identity, which is roughly seven times what the model can resolve. Then the motor warms up: resistance climbs while magnet strength falls, both pushing the same way, and at 34 °C the drift exceeds the entire difference between units. A warm motor reads like a different cold one, with nobody attacking anything. Switch the thermometer on and it comes back. Temperature is solved; wear is not, and that is the measurement still waiting to be made.
The instrument reports a null control beside every margin, and says no answer available for the industrial arm rather than guessing at one. There is exactly one published method for measuring robot energy, ISO/TS 25213, effective 26 June 2026. It covers six-axis industrial arms, it is paywalled, and being a voluntary specification it obliges nobody. Autonomous mobile robots are excluded from it by name and humanoids are not mentioned at all, which is to say the bodies Physical AI actually uses have no method yet. That is the opening rather than the obstacle: an open method, free to run, stating its plane, its noise floor against a known load, its null control and its residual, is more adoptable than a paywalled one.