No organism observes its own full state jointly. A nervous system holds local models at local nodes, each fit to what crosses that node, and nothing anywhere holds the global state vector. The causal abstraction mathematics this course leans on asks when such local models cohere into one global account, and answers with a condition on how the models fit together, not on how much data any of them saw. Here is that theorem's smallest possible demonstration. Three correlations, each estimated perfectly, each a completely valid model of its own pair. Assemble them and ask for the joint distribution. A correlation matrix must be positive semidefinite to describe anything that can exist; this one has an eigenvalue of minus 0.8. There is no joint model. Not hard to find: nonexistent. And because each local piece is already exact, more samples of any pair cannot help. The failure is in the loop the three pairs form, and it switches on at a sharp threshold you will measure. Then the same question, counted on a real machine: four teams share one dual-arm robot carrying a rod, each team writes its own body's equation of motion, and the count says the torso can resolve one of its four port unknowns, each arm four of five, the payload three of six, and the whole robot all ten. Every team is short. The machine is not.