We propose a formal framework for measuring the extent to which the structural features of a world are knowable by an observer internal to that world. We introduce the Elementary Predictive Step (EPS, ε), a relation-property of pairs (observer, regularity), defined through the observer's predictive performance under a specified model class. We construct an operational scale of a world relative to a model class and encoding convention, and define a knowability coefficient η as the mean of ε values over this scale. We distinguish between deep closure (a model achieved through the observer's own prediction-validation loop) and transmitted closure (a model acquired from another source), and discuss properties of each mode that may differ under conditions for which the framework prepares investigation. The framework synthesizes elements from educational data mining, the predictive processing tradition, structural realism, algorithmic information theory, and the epistemic horizons literature, with the aim of providing definitions sharp enough to support replicable empirical work on the limits of internal cognition in simulated worlds. This paper introduces no empirical results; it prepares the apparatus for the empirical papers that follow.
Ivan Denysov (Tue,) studied this question.