One steady sideways push, one and a half newtons, blowing through training and test alike. No distribution shift, no trick. The spring now holds a displacement all episode long, the way your arm holds gravity all day, and holding is work with a direction. The observer books that holding force to the motor, so the reconstructed labels now carry a standing bias where before they carried a doubled restoring force. The true-label clone keeps its perfect score. The reconstructed one scores zero, and look at how it fails: every fall lands in a three-tenths window around four seconds, exactly as the target swings back through centre. Noise does not fail on a clock. A structural error does, when the task finally asks its question. Map the toy home: hand-pose reconstruction on load-bearing contact is this arm of this bench at scale, and carrying, pressing and holding are where Physical AI most needs to work. The fix costs a logging change, not a fleet: record the command, declare the substrate, keep reconstruction for the parts of the loop that do no directional work.