Ask five people in active inference what a Markov blanket is, and you will get five sincere answers that quietly disagree. That disagreement is not a scandal. It is the honest state of the concept, and worth pulling apart before you use it to build anything.
The definition, kept minimal
In its original probabilistic-graphical-models sense (Class E), a Markov blanket of a node X is the smallest set of other nodes that renders X conditionally independent of everything else in the graph. Once you know the blanket, adding any variable outside it tells you nothing new about X. That is the entire mathematical content. It is a statement about conditional independence in a joint distribution, nothing more.
Parr, Pezzulo, and Friston (2022) carry this into a partition of a system and its environment (Class E). States get split into four sets: internal states, external states, and two blanket sets that mediate between them. The blanket sets are called sensory states (what the internal side reads from the outside) and active states (what the internal side writes to the outside). The claim is that internal and external states are conditionally independent given the blanket.
Sensory and active states, briefly
Sensory states are the observations the internal side receives. They depend on external states plus noise, and they are what the generative model has to explain. Active states are the outputs the internal side emits back into the world. They depend on internal beliefs and, through the environment, help shape the next round of sensory input. Perception updates beliefs to explain sensory states. Action updates the world so that future sensory states become the ones the model expected. Both reduce variational free energy, the same quantity, through different routes (Class E).
The elegance is real. So is the risk of over-reading it.
What the blanket is not
A blanket is not a physical membrane. It is not skin, it is not a cell wall, it is not an API boundary carved into hardware. It is a statistical cut through a joint distribution, valid in the model where it was defined. Two modelers who partition the same physical system differently can draw different blankets, and both can be internally consistent (Class G).
A blanket is not, by itself, evidence of a "self". The presence of a conditional-independence structure does not tell you that anyone is home. Recent literature has pressed on this exact point: several authors argue that the identification of a blanket with an agentive self is stronger than the mathematics supports, and that in continuous-time stochastic systems the blanket assumption can be difficult or impossible to hold strictly (Class E, Class F). The debate is open. We take that as a working caution, not a punchline.
A blanket is not universally stable. In many real systems the set of variables that renders internal and external states independent changes over time, because the system itself changes what it senses, what it does, and what counts as "inside" (Class C, Class U). If you fix the blanket at t=0 and let the system run, you are asserting more than the formalism gives you.
The debate, in one honest paragraph
There is a live disagreement in the literature about how much explanatory weight the blanket can carry. One camp treats blanket structure as a derivable property of certain classes of stochastic differential equations, and argues that self-organizing systems tend to acquire it. Another camp argues that this move quietly imports the conclusion into the premises, and that in general dynamical systems the blanket assumption has to be posted, not proved (Class E, Class F). Both camps have serious papers. We treat the concept as useful, and we do not treat it as settled. That is our public position, and it is the same position we bring inside the workbench.
How UNI uses blanket structure
Inside UNI the blanket is a scoping device. When we set up a POMDP in the Precision Lab, the agent has a generative model over hidden states, an observation model that maps hidden states to sensory readings, and a transition model that maps hidden states plus actions to next hidden states (Class B). The variables named "sensory" and "action" are the blanket sets in that model. That is where the blanket lives for us: in the graph we wrote down, in the code you can read.
What our implementation actually enforces:
- Sensory access is scoped. The agent reads only what the observation model exposes; it never queries the hidden state directly (Class B). This is checked in the falsifier suite for the Cell Lab: an agent that can peek scores differently, and that is the disconfirmation criterion (Class F).
- Action writes only through the transition model. The agent does not mutate the environment except through the action channel (Class B). Any drift from that would show up as an integrity failure in the run log.
- Internal beliefs are recomputed from the blanket alone. The posterior over hidden states is a function of the sequence of sensory readings and actions, not of the true environment (Class B).
What our implementation does not claim:
- We do not claim the blanket "emerges" from the code. It is written into the code (Class C). Any claim otherwise would be a category error we are happy to be called on.
- We do not claim the blanket certifies an inside self. The agent is a scope, not a subject (Class U).
- We do not claim the same blanket holds across labs. Each lab draws its own partition; comparing across them is a modeling exercise, not a fact (Class U).
Where this fits in the family
For a builder's map of the surrounding vocabulary, see Active Inference Fundamentals, A Working Map. For the distinction between hidden state and observation, which is where the blanket does its heaviest lifting, see Hidden States vs Observations, A Builder's View. For the split we actually enforce, see Sensory and Active States, A Working Decomposition. For our public stance on the interpretation debate, see The Blanket Debate, What We Take As Working Assumption. For how we test whether any of this is actually earning its keep, see Gates and Falsifiers, How We Know When We Are Wrong.
A note on Themesis
Themesis maintains a public resource map for people entering active inference in 2026. It lists SolutionWright among several starting pathways, alongside independent educational and research groups. We link it here as a factual pointer for readers who want a wider view of the field, not as an endorsement of our specific claims: Where to Start with Active Inference, A Resource Map for 2026. Our one-line frame, in our voice: the map that lists SWU among five pathways; anchor for family tie-ins.
What the reader should take away
Treat the Markov blanket as a useful scaffolding you place over a joint model, not a discovery about the world. It buys you a clean way to say "here is what the agent can see, here is what the agent can do, and here is the belief loop that connects them." It does not buy you selfhood, autonomy, or life. If someone claims it does, ask which papers they are leaning on, and read the counter-papers.
In UNI, the blanket is a load-bearing modeling choice we take seriously and check publicly. When we get it wrong, the falsifier suite is meant to catch it. That is the deal we are offering.