Intervention-Based Identification of Self-State Representations

Can a system figure out which of its own internal representations is about itself, using only its own actions? This note reports an experiment that tests exactly that question, inside a small p-adic algebraic system. The system is given two internal variables with identical structure. One is coupled to the system's own state; the other is coupled to the environment. The system has no built-in information about which is which. The driver of the state equation is chosen at random for each run. The system applies a controlled intervention to itself and observes which variable responds. Across 20 randomized assignments, the coupled variable diverges by 15 while the uncoupled one stays at 0 — a separation set by the causal structure of the dynamics, not by noise. The identified variable is then used for control. Relying on it, the system holds a target invariant for 99.8% of steps; relying on the wrong variable, viability drops to 0.3%. A complete corruption sweep closes the experiment: shifting only the copy of the representation supplied to the policy moves the closed-loop fixed point by exactly 2δ modulo 13, and the observed viability map matches the predicted modular tolerance set for all 13 values of δ. What the experiment demonstrates is a specific, operational capability: a system can intervene on its own dynamics, identify which of two structurally symmetric internal representations is coupled to its state, and use the identified representation in a content-sensitive feedback loop. The test protocol is portable to any architecture meeting the same conditions. The experiment uses the Zeta system (Dávid Navrátil) as substrate. The hidden-state dynamics, probe, and control policy are additions made for this test. The note does not claim consciousness, semantic selfhood, robustness to noise, or scalability. Code is included.