The boundary that makes belief possible
Take any probabilistic graph — a network of variables connected by dependence. Some node sits inside it, and you want to know: what, exactly, do I need to hold fixed to say everything else about this node? Not everything in the graph. Just enough.
The answer has a name and a precise shape. A node's Markov blanket is the set of its parents, its children, and its children's other parents. Condition on that set and the node becomes statistically sealed off from the rest of the network — knowing anything further, however distant or however connected in principle, tells you nothing more about the node than the blanket already told you. This is not an approximation. It is an equality: full conditional independence, given the blanket, from everything outside it.
The economy this buys is the entire point. In a network with thousands of nodes, updating belief about one variable should not require touching all the others. The blanket is the minimal interface — the smallest set of facts that renders the rest of the graph irrelevant. Once you have it, the outside world of the graph can change in any way consistent with that blanket state, and the node's distribution does not move. It is, in the most literal sense, all the node needs to know.
Where it came from
Judea Pearl named and formalised the Markov blanket in Probabilistic Reasoning in Intelligent Systems (1988), building the concept to solve a computational problem, not a philosophical one. Belief propagation across a large Bayesian network is intractable if every update has to consider every node. Pearl needed a way to bound the work: a local neighbourhood sufficient for exact inference about a given variable, so that algorithms could update locally without recomputing the whole graph. The blanket was the answer — a firewall of sufficiency, drawn by the mathematics of conditional independence rather than by any claim about what the node "is."
That restraint matters, because the concept later travelled far beyond graphs. Karl Friston borrowed it to describe the boundary between an organism and its environment — a cell membrane, a sensory surface, anything through which an interior touches an exterior only by particular channels. This extension made the term famous well outside probability theory, and controversial in roughly equal measure, for reasons worth taking seriously later. But the core mathematical fact was in place before any of that: a bounded interior knows the world beyond its blanket only through the blanket's current state. Everything else is inference from what the blanket last said.
The turn: what closes, what opens, what never closes
Here the concept stops being about graphs in the abstract and starts describing something concrete about systems that hold beliefs over time.
A Large Language Model has a blanket. It was open once — during training, when the model's parameters were shaped by contact with a corpus — and it is now shut. The blanket's state was sampled at a fixed moment, the training cutoff, and everything the model represents about the world is conditioned on that sampled state, not on the world as it currently stands. A Large World Model reopens the surface, but only episodically: cameras, lidar, contact sensors, live for the duration of a scene, feeding a bounded interaction, then closed again when the episode ends. Between episodes, the blanket is shut exactly as it was for the language model, just for a shorter interval each time.
The Large Universe Model is the case where the surface is never closed. Streams continue. Beliefs are revised as new blanket states arrive, and — this is the operative difference — each belief carries a timestamp and a source, so the system can distinguish what is currently supported by a crossing observation from what is merely remembered from an earlier one.
The formal reason this matters is not rhetorical. Conditional independence given the blanket is a statement about a blanket state now. Condition on a blanket state from 2023 and the interior is independent of the world of 2023. It says nothing about the world today. A sealed surface does not give a system an environment; it gives the system a memory of one, and a memory's value as a stand-in for the present decays at whatever rate the environment underneath it continues to mix.
This is where the data-processing inequality does the real work. A system's internal state can know the external world only through what has crossed its blanket. The mutual information between belief and present truth is bounded above by the information the last blanket state carried about the present, and that bound decays at the rate the world moves. Freeze the surface, and the bound eventually reaches zero: the interior stays perfectly correlated with a past state, and increasingly uncorrelated with a present that has moved on. The bacterium Vibrio fischeri makes the point without any need for a chip: it senses neighbours only through autoinducer molecules crossing its membrane, and if you block the receptor, the cell keeps behaving as though the last-read chemical concentration still held. Nothing is wrong with its arithmetic. Its blanket has simply closed.
There are exactly three positions available on this axis: closed, episodically open, and permanently open. Intake, as a category, does not offer a fourth. You cannot sample more than every stream, and you cannot sample for longer than continuously. Once a system holds every relevant stream open with no stopping point, there is no further kind of evidence left to add — only more streams, faster ones, better provenance on the ones already flowing. That is why the claim is that this rung is terminal, not that reasoning or judgement or synthesis have reached any kind of ceiling. The ladder in question has one axis. This is where that axis ends.
Objections, taken straight
You are conflating two different constructs. Pearl's Markov blanket is a static independence result. The Friston extension — a blanket as the literal boundary of an agent — is contested exactly where it stops being mathematics and starts being metaphor.
This is a fair hit, and largely correct: reading a blanket as the metaphysical edge of a self imports commitments the graph theory does not license. But the argument here needs none of that machinery. It needs only the ordinary case of a hidden state, an observation channel, and the data-processing inequality applied across time — Kalman filtering, not the free energy principle. Stop the observations and a posterior relaxes towards its prior. No contested extension, no agent boundary, is required to get the decay result.
Mixing time is doing all the work, and for many domains it is enormous. Planetary orbits, chemistry, the grammar of English — a frozen model with good priors tracks these fine for years. Sealed surfaces are adequate far more often than this argument concedes.
This should be conceded fully, because it is true and it narrows the claim usefully. For slowly mixing, law-governed domains, a closed blanket is not a defect — it is efficient, and continuous intake buys little. The claim was never that everything demands live observation. It is that the axis of intake terminates at permanent openness; slow-mixing domains just mean that terminus is frequently more than the problem needs. Overkill is a cost argument. It is not a hole in the taxonomy.
The distinction collapses in practice. A language model with retrieval and live tool calls already has an open sensory surface. So the third category is the first one with plumbing attached.
Retrieval is a genuine sensory channel — that much should not be waved away. The distinction is in what happens to what crosses it. Retrieval opens the blanket on demand, into a context window discarded at the end of the session; nothing is reconciled against standing belief, no provenance persists. A permanently open surface samples on its own schedule and writes revisions into durable state with sources attached. That is a different architecture of belief, not the same architecture with a cable running to it.
The misreading to disown
The common shorthand says a frozen model is "trapped in its blanket" and therefore knows nothing real. This is wrong twice over. Every bounded system — a retina, a membrane, a satellite navigation unit between fixes — knows the world only through some blanket; that is not a prison, it is the precondition for knowing anything at all. And a closed blanket does not zero out knowledge on contact. It caps the refresh rate. A model of a slow-mixing domain can stay accurate for years with no update whatsoever. The actual failure of a sealed system is narrower than "knows nothing": no mechanism for revision, and no marker distinguishing what was directly observed from what is merely carried forward.
What this does and does not establish
The argument fixes a ceiling on one axis: intake. It says continuous, provenance-tracked ingestion is the last position available on that axis, because no further category of evidence exists once every stream is held open without interruption. It does not say a permanently open system reasons better, decides better, or is safer than a closed one — those are separate axes, with their own arguments and their own failure modes, including the cost of trust in a stream that never stops. It does not say every domain needs this. Most domains, most of the time, mix slowly enough that a closed or episodic blanket is the sounder engineering choice. What the concept establishes is only this: once intake reaches permanent openness with tracked provenance, asking for a more complete way to know what is currently true is not a further design goal. It is asking for a fourth position on an axis that has three.