We encounter a system only through its responses, yet the system's state refers to its intrinsic condition. If we can predict future responses, we may acquire knowledge beyond what is directly observed. Here we give “state” a model‑relative meaning: it is the complete response profile over a specified query universe, and any two configurations that agree on every query are considered the same state.\ \ Within this view, a learned state is understood as the set of distinctions it retains among different response profiles. Prediction deals with a selected subset of observations, whereas recovery concerns the entire profile. We explore their relationship by examining query ambiguity, the geometry of the state space, and the choice of prediction targets.\ \ Our investigation uses two complementary settings: synthetic systems with known response laws, providing a controllable ground truth, and neural recording data, illustrating application to real biological systems. The results show that when prediction targets span a sufficiently rich set of queries, the act of prediction itself can progressively reconstruct the full response profile, thereby achieving ontic state recovery.\ \ The aim of this work is to clarify how prediction can serve as a way of knowing a system, and what it means for a learned state to preserve the complete information of the studied system.\ \ Review