Do Markov blankets have any ontological weight? Most of the literature on the free energy principle (FEP) equates Markov blankets with actual system boundaries without further ado. However, as the authors rightly point out, there is a crucial difference between using blankets as an epistemic tool in Bayesian networks to identify independencies among random variables (Pearl blankets) and using them as actual boundaries between a system and its environment (Friston blankets).

I agree with the authors on the insufficiency of Friston blankets to demarcate the physical boundary between an agent and its environment, and the need for additional philosophical assumptions “to do such heavy metaphysical lifting.” FEP theorists seem aware of this since they recently gave their framework a freshly minted metaphysical interpretation, dubbed Markovian monism (MM) (Friston, Wiese, & Hobson, Reference Friston, Wiese and Hobson2020). According to MM, the very fact that one may demarcate a system from its environment through a Markov blanket induces a dual aspect information geometry in the system's internal states that enables it to represent its surroundings. MM thus reveals the troublesome transition between Markov blankets as epistemic model-bound tools and their alleged ontological consequences, which are beset by circular reasoning (Sánchez-Cañizares, Reference Sánchez-Cañizares2021).

However, within FEP formalism, the ontology proper to Markov blankets need not take the brunt of the blow because it inherits the more general problem of distinguishing between a system and its environment in physics. In the absence of a theory of everything, conventional physics must accept initial conditions to start computing the world's behavior. “You need a starting point!” (Wilczek, Reference Wilczek2015). Remarkably, the existence of systems seems to require the universe to have exceptional initial conditions and dynamics (Tegmark, Reference Tegmark2015) since the theory that is currently most fundamental, that is, quantum mechanics, deems the distinction between system and environment as relative (Lombardi, Fortín, & Castagnino, Reference Lombardi, Fortín, Castagnino, de Regt, Hartmann and Okasha2012). Physics teaches us that one must make an ansatz to progress in the scientific description of nature. Whether a chosen ansatz holds at a specific process only becomes evident a posteriori. Markov blankets are undoubtedly the most basic ansatz for FEP formalism to work.

Things being so, the authors' critiques of Markov blankets as Friston blankets become less weighty in two interrelated respects:

1. (1) Markov blankets define the dominion – via the partition of variables – for which it makes sense to minimize free energy. Their precise definition must change whenever the model becomes unsuitable for describing unexpected dynamics, that is, dynamics that cannot be fully grasped within the assumptions of a particular model. But this does not differ from the usual procedure of changing boundary conditions for the distinct models that are compatible with a principle theory. The authors themselves recognize this when quoting Andy Clark, “boundaries are malleable (over time) and multiple.” FEP is a principle and needs reinterpretation for each model. One such reinterpretation must state what the system and its environment are, even though such a distinction does not strictly stem from FEP formalism. As a consequence, one should not consider FEP as wholly explanatory of living beings' natural history (Longo & Montévil, Reference Longo and Montévil2014), if only because Friston blankets also change throughout the system's history.

2. (2) Even though Friston blankets are metaphysically wanting, they should not be judged as a mere instrumental tool. The authors, however, ultimately lean toward “a strongly instrumentalist understanding of Bayesian networks, and hence of Markov blankets, which would not justify the kinds of strong philosophical conclusions drawn by some from the idea of a Friston blanket.” Instrumentalism undoubtedly looms large in philosophical interpretations of scientific research, but, in the end, this seems to be a self-defeating strategy for several reasons:

   1. (2.1) The instrumentalist that deems Markov blankets as Pearl blankets forgoes endowing the former with any ontological weight. Yet such a mindset could easily lead to denying the very existence of systems – as sheer constructs of human perception. Nevertheless, if one does not wish to embrace such a radical position and, on the contrary, accepts the existence of systems in the universe, something quite similar to Friston blankets must also exist in each system in order to sieve through the many environmental influences that foster or threaten the system's identity.

   2. (2.2) It seems paradoxical to emphasize the insufficient justification for transforming Pearl blankets into Friston blankets whereas, ultimately, glossing over what additional philosophical assumptions might look like for such a move. A scientific realist, for instance, could call on formal causation as a valid metaphysical framework that allows for the use of ad hoc boundary conditions to individuate systems that enjoy specific dynamics in nature (Owen, Reference Owen2020, Reference Owen2021; Sánchez-Cañizares, Reference Sánchez-Cañizares2022a, Reference Sánchez-Cañizares2022b). In other words, Markov blankets reflect the emergence of boundary conditions for living systems. If boundary conditions exist for some relevant time scale, Markov blankets are Friston blankets. In addition, such a philosophical commitment can adequately frame the emergence of complex dynamical systems, which one cannot just deduce from their underlying dynamics (Juarrero, Reference Juarrero2002; Sánchez-Cañizares, Reference Sánchez-Cañizares2016).

   3. (2.3) The methodological issue at play refers to whether other kinds of knowledge – for example, knowing living systems as wholes – that influence scientific research should be accepted within the overall explanatory picture. Certain pre-scientific knowledge is necessary for guiding scientific methodology. If one assumes said knowledge, there is no fundamental reason to deny that some Markov blankets are also Friston blankets, even if for a limited period, or that Friston blankets are not fixed and may transition in a variety of ways toward different instantiations. In doing so, the inevitable, closed circularity of scientific instrumentalism turns into the open circularity of scientific realism, which admits a hierarchical variety of assumptions and hypotheses about the world, as well as the possibility of cognitive progress based on constant confrontation with observation.