The scientific process plays out in a multi-scale system comprising subsystems, each with their own properties and dynamics. For the practice of science to generate useful world models — and lead to the development of enabling technologies — practicing scientists, their theories, methods, dissemination, and infrastructure (e.g., funding and laboratories) must all fit together in an orchestrated manner. Scientific practice has broad societal implications that go beyond mere scientific progress: we base our decisions on theoretical (i.e., models and forecasts) and technological (e.g., vaccines and smartphones) scientific advances. This paper applies the free energy principle to provide a multi-scale description of science understood as evidence-seeking processes in a nested hierarchy of living (biological and behavioural) and epistemic (linguistic) structures. This allows us to naturalise the scientific process — as distributed self-evidencing — in terms of dynamics that can be read as inference or Bayesian belief updating; i.e., processes that maximize the evidence for a generative model of the sensed and measured world. The ensuing meta-theoretical approach dispels the notion of science as truth-pointing and foregrounds inference to the best explanation — as evinced by the beliefs of scientists and their encultured niche. Crucially, it furnishes a way of simulating the practice of science, which may have a foundational role in the next generation of augmented intelligence systems. Epistemologically, it also addresses some key questions; e.g., is science a special? And in what ways is scientific pursuit an existential imperative for all beings? These questions may be foundational in how we use and design intelligent systems.