Process models and event logs are the two most important process representations in process mining. When comparing the two, e.g., to reveal discrepancies between them, we can consider either the model-log perspective, meaning the degree to which the model represents the event log, or the model-system perspective, meaning the degree to which the model represents the underlying process. The two perspectives complement each other, but so far, most research has focused on assessing model-log similarity. In this paper, we propose a novel framework for assessing model-system similarity that addresses the two major challenges that this task poses. First, it enables a comparison between model and system that is independent from the concrete modeling formalism by measuring their similarity based on the n-grams of their respective languages. Second, it abstracts from the event log as an incomplete sample of the system by projecting the likely language of the system by means of statistical estimators. Our empirical evaluation shows that this framework provides a valid way to assess model-system similarity, which can be used in many different applications.