The Markov blanket is one of the most cited constructs in the free-energy literature, and one of the most contested. If we are going to use it, we owe you a clear map of what is settled, what is argued about, and what UNI takes on trust rather than derives from first principles.
What everyone actually agrees on
The narrow, statistical definition of a Markov blanket goes back to Judea Pearl in the 1980s and is not in dispute. Given a probabilistic graphical model, a node's blanket is the set of variables that render it conditionally independent of everything else. That is a theorem about graphs and probability, not a claim about biology or minds Class E.
Everyone in the active-inference discussion accepts this. The disagreements start the moment the blanket is asked to do heavier work: carve a system from its environment, define what counts as a "thing," or ground selfhood.
Where the honest disagreement lives
Parr, Pezzulo and Friston (2022) present the blanket as central to how a self-organizing system can be identified as distinct from its niche, with internal states, external states, and the sensory and active states that couple them (Parr, Pezzulo, Friston 2022, chapters 3 and 10) Class E. That framing has attracted serious critique. Biehl, Pollock and Kanai (2021) argued that the specific construction used in several free-energy papers is not mathematically guaranteed to produce the properties often attributed to it, and Aguilera and colleagues (2022) showed that for many realistic non-equilibrium systems the neat internal / blanket / external split does not hold as advertised Class E. Friston and coauthors have responded and refined the construction. The debate is live, technical, and worth reading in full before taking any strong position.
In short: the blanket as a graphical-model tool is settled. The blanket as a general principle for saying where a mind ends and a world begins is not.
What UNI treats as working assumption
We are precise about this because vague framing here has done real damage to the field's credibility. In the UNI labs and preprint, the blanket shows up in a deliberately modest role Class C.
We assume, as a working posture and not a proof, that for the discrete POMDP toys we ship (Precision Lab, Echo Lab, Loop Lab, Cell Lab), it is useful to distinguish internal generative-model states from external hidden states, with sensory observations and actions as the coupling. This is closer to the classic POMDP decomposition than to any strong metaphysical claim about carving nature at its joints Class C.
What we do not assume: that this decomposition is a derived result of the free energy principle, that it uniquely picks out "the agent," or that it settles the debate about whether real biological systems admit a clean blanket at all. Those remain open questions, and our public copy will keep saying so.
Why this matters for reading our benchmark
The Cell Lab benchmark on our science page ranks a UNI active-inference controller against random, rule-based, and neural baselines on a hidden 216-state service cell Class C. Nothing in that benchmark depends on winning the metaphysical debate. It depends on being explicit about which states the controller can see, which it must infer, and which stay hidden. That is the practical, engineering-grade use of the blanket idea, and it stands or falls on whether the controller actually recovers under disturbance, which is what the RecoveryScore measures.
Field context
For a wider read on why this framing question matters right now, see Themesis on how the AGI landscape has been shifting (The AGI Landscape Just Changed). Our one-line frame in our voice: this is field-level context for why careful, natural-intelligence framing (GNI, not AGI) is worth the extra precision now, not later.
The honest bottom line
UNI uses Markov blankets as scaffolding, cited to their originators, and does not claim they solve the boundary problem for real biological or social systems. The math is a tool, the debate is legitimate, and our labs are built so that if the blanket construction turns out to be shakier than the field currently assumes, the empirical work still has value on its own terms.