Abstract: This will be a discussion of some recent work in progress, not a talk. Current machine learning and statistical practice relies on the particular assumption that an individual's features are intrinsic properties separated from the processes that summon forth the featurized individual in the first place. Working under this paradigm makes analysis amenable, but hinders accuracy in its effective goal of modeling real world systems. A notable example can be found in the recent research interest of "fairness" in machine learning, in which discrimination based on certain categories (race, gender, class, etc.) is intended to be avoided. The limitations surface when these categories are assumed ontologically independent of other ascribed properties of the individual, and pre-existent to the same mechanism which is analyzing or making inference on them. To deconvolve these assumptions and reformulate our practice, we will take race as an example and discuss ways in which racial categories, understood from different theories of race, can be better described as mathematical "structures of the system", in opposition to atomic "properties of the individual". We will discuss the benefits, challenges, and potential of these views for non harmful techne of category-based machine learning and causal inference.