NA-04 — MIND / BODY / MIND.BODY / WORLD — the Markov blanket, nested across scales

What you are reading. The formal object behind the phrase "mind body mind.body world": the Markov blanket. This chapter gives the partition exactly, maps each of the four words onto one of its four sets, quantifies a real blanket with measured anatomy, walks the nesting that produces the gradient from organelle to ecosystem — and then spends its last third arguing against itself, because the commonest error in this literature is treating a blanket as a fact you may assume rather than a modeling choice you must justify. A blanket is a modeling choice. Everything downstream of that sentence is the chapter.

---

Anchor: this extends M12, it does not rival it

This corpus already carries the map, as kitchen rule M12 — WORLD ⊥ BODY ⊥ MIND (cookbook/01-kitchen-rules.md): "Two typed Markov blankets; interoception = hardware signals," with a discrete POMDP perceive → EFE-plan → act → learn and textbook-level F[q] ≥ −ln p(o|m). M12 is status: method — a governance pattern, not a capability claim and not an observation about nature. This chapter supplies the natural-science reading of the same partition and raises nothing: M12's status is unchanged by every citation below.

The operator's fourth term, MIND.BODY, is the one M12 leaves implicit. It names the interface — and it is the load-bearing part.

---

The partition, formally

Partition a system's states z into four disjoint subsets: internal μ (MIND), sensory s, active a, external η (WORLD). The blanket is b = {s, a} — BODY. The defining condition is a conditional independence, and it is the whole content of the idea:

p(μ, η | b) = p(μ | b) · p(η | b)          ⟺   μ ⊥ η | b   ⟺   I(μ ; η | b) = 0

In words: given the blanket, internal and external states carry no further information about each other. Every dependency between mind and world is routed through the body. Not mostly. Not usually. By definition — or the partition is not a blanket.

The term is Pearl's, coined in a graphical-model setting: Pearl (1988), Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference, Morgan Kaufmann — the minimal set of variables rendering a target conditionally independent of all others. It was a tool for efficient inference over a graph you already drew. Hold onto that; it is the pivot of the fence below.

The free-energy literature adds a directional sparsity requirement on top: the canonical perception–action partition, in which sensory states are influenced by external but not internal states, active states by internal but not external states, and the mind–world flow blocks vanish:

∂μ̇/∂η = 0        and        ∂η̇/∂μ = 0

For linear Gaussian systems this reduces to the vanishing of the internal–external blocks of the precision (inverse-covariance) matrix, H_μη = H_ημ = 0 — the form in which the condition was solved exactly and, as the fence records, largely failed. See Friston (2013), "Life as we know it", J R Soc Interface 10(86):20130475, DOI 10.1098/rsif.2013.0475; Friston (2019), "A free energy principle for a particular physics", arXiv:1906.10184.

---

MIND.BODY — the interface is the thing

Read the partition again and notice what it forbids. μ has no term in η. The mind's only evidence about the world is s; its only leverage on the world is a. Therefore:

A mind can never touch the world. It can only ever touch its own blanket.

Every inference is an inference from sensory states; every act is a perturbation of active states. The world is reached exactly never — only ever inferred through a surface. The generative model is not a model of η; it is a model of how η is expected to show up in b. The world enters as a hypothesis that explains the blanket.

This is why embodiment is not decoration. The blanket is not a wrapper around the interesting part — it determines what is inferable at all. Change the surface and you change the set of possible minds behind it, because you have changed the only evidence there will ever be. A system with two sensory channels and one with seven do not have the same world available, however good the inference machinery. MIND.BODY is not a fifth thing: it is the coupling μ ↔ b, a pair of one-way doors — s the world's only route in, a the mind's only route out.

The blanket, quantified

A blanket is a real object with a real width, and the human visual blanket has been counted. It is a funnel: 92 × 10⁶ rods and 4.6 × 10⁶ cones on the transducing face (Curcio et al., 1990) against ≈1.16 × 10⁶ optic nerve axons carrying it out of the eye (Jonas et al., 1990) — roughly ~83 photoreceptors per outgoing axon (a MODELED ratio: it divides numbers from different cohorts; see the table). And the surface is radically non-uniform: peak foveal cone density averages 199,000 cones/mm², ranging 100,000–324,000 between individuals (Curcio et al., 1990). The blanket is not a window. It is a lossy, unevenly weighted, individually variable compression stage — and it is the entire visual world's only door.

---

Interoception: a second blanket inside the first

The body also senses itself. Cardiac, gastric, respiratory, inflammatory and osmotic channels are sensory states whose external states are other parts of the same organism — a blanket nested inside a blanket. From the brain's position, the viscera are η.

Two groups arrived at the predictive reading independently: Seth & Friston (2016), "Active interoceptive inference and the emotional brain", Phil Trans R Soc B 371(1708):20160007, DOI 10.1098/rstb.2016.0007 — autonomic reflexes enslaved by descending predictions; and Barrett & Simmons (2015), "Interoceptive predictions in the brain", Nat Rev Neurosci 16(7):419–429, DOI 10.1038/nrn3950 — the EPIC model, in which agranular visceromotor cortices issue interoceptive predictions rather than receive interoceptive reports. This corpus's own version is M12's blunt engineering clause, "interoception = hardware signals." Right shape, status: method; the biology above does not raise it.

The interoceptive channel is also where this chapter earns its contested row. The famous figure — "~80% of the vagus is afferent" — is real, replicated, and routinely quoted outside its scope: it is Prechtl & Powley (1990), measuring the rat abdominal vagus. The human cervical vagus, by immunofluorescence in eight cadavers, gives sensory fractions of 73.9 ± 7.5% (right) and 72.4 ± 5.6% (left), with ~13% parasympathetic and ~13% sympathetic fibers riding along (Kronsteiner et al., 2024). And the total count is disputed by ~4× between methods: ~100,000 axons classically (Hoffman & Schnitzlein, 1961) versus ~23,000–25,500 modern (Kronsteiner et al., 2024). Carry both. The dispute is the finding.

---

Nesting: blankets of blankets (the gradient)

Nothing in the definition fixes a scale. A blanket's internal states can themselves be a set of blanketed things, and the construction recurses:

organelle → cell → tissue → organ → organism → colony → ecosystem

Each level's blanket becomes part of the next level's internal states. The formal treatment is Palacios, Razi, Parr, Kirchhoff & Friston (2020), "On Markov blankets and hierarchical self-organisation", J Theor Biol 486:110089, DOI 10.1016/j.jtbi.2019.110089 — "(macroscopic) Markov blankets of (microscopic) Markov blankets." See also Kirchhoff, Parr, Palacios, Friston & Kiverstein (2018), "The Markov blankets of life", J R Soc Interface 15(138):20170792.

This recursion is the gradient — the formal reason one vocabulary addresses a cell and an ant colony without changing. The full ladder with its measured scale constants is NA-10; not duplicated here. Two honest notes, immediately:

1. The recursion is a construction, not an observation. That blankets can compose this way is a mathematical fact. That any particular biological hierarchy is so composed is a separate empirical claim — and the next section is about how rarely it has been checked.
2. NOT-MEASURED: the number of nested blanket levels in any named organism, established by measurement rather than by choosing a diagram. There is no such count to report.

---

When is a new blanket real? (the operable criterion)

A candidate boundary b around a candidate interior μ is a real blanket exactly when conditional independence actually holds across it:

I(μ ; η | b) = 0        (nats; estimated over the system's actual trajectories)

How you would measure it. Estimate the conditional mutual information between proposed interior and proposed exterior, conditioned on the proposed boundary, from observed dynamics. I ≈ 0 within estimator noise ⟹ the boundary does the work claimed. I > 0 ⟹ there is a route from world to mind that bypasses the body, and the partition is wrong — the boundary is drawn in the wrong place. Constraint-based blanket-discovery algorithms in machine learning (IAMB, GS, HITON-MB) do exactly this kind of conditional-independence testing to find blankets rather than assume them.

And now the honest part: this is almost never done. Conditional-independence testing is notoriously hard for continuous, high-dimensional, non-stationary variables — which is to say, for every interesting biological system. The attempts are mostly formal or simulated. Friston et al. (2021), "Parcels and particles: Markov blankets in the brain", Network Neuroscience 5(1):211–251 (arXiv:2007.09704), gives a renormalisation-group treatment of blanket partitions over effective connectivity — whether that detects a blanket or imposes one is precisely what critics dispute. Beck & Ramstead (2025), "Dynamic Markov Blanket Detection for Macroscopic Physics Discovery" (arXiv:2502.21217), builds a variational-Bayes detection algorithm and demonstrates it on Newton's cradle, a burning fuse, the Lorenz attractor, and simulated cells — their own framing is "simple numerical experiments." Not a mouse. Not a mitochondrion.

The criterion exists, is clean, and is overwhelmingly unexercised. When a paper — or this corpus — says "system X has a Markov blanket," the default assumption should be that I(μ;η|b) was never estimated.

---

Action as inference

The partition makes prediction error resolvable two ways, and this is the part that most repays attention. Given a mismatch between predicted and actual sensory states, a system can change μ — update the model to fit the world — or change a — change the world until it fits the model. Both minimize the same quantity. Action is not a separate faculty bolted onto inference; it is inference run through the other half of the blanket.

A real example. Adams, Shipp & Friston (2013), "Predictions not commands: active inference in the motor system", Brain Struct Funct 218:611–643, DOI 10.1007/s00429-012-0475-5, argue that descending motor signals are proprioceptive predictions, not motor commands, and that classical spinal reflex arcs discharge them: the cord receives a prediction of where the limb will be, finds a mismatch with where it is, and resolves it by moving the limb — the arc fulfills the prediction rather than reporting the error upward. Their worked case is the knee-jerk reflex. Whether this is the correct account is unsettled; the table cards it HYPOTHESIZED, not higher.

Thermoregulation makes the symmetry plainest: a cold mammal can revise its prediction of its own temperature, or it can shiver. Only one is survivable. The preference C over sensory states breaks the tie — which is why the free-energy story needs preferences and cannot be pure inference.

---

The honest fence: Pearl blankets vs Friston blankets

This is the most important section here, and it cuts against everything above.

Bruineberg, Dołęga, Dewhurst & Baltieri (2022), "The Emperor's New Markov Blankets", Behavioral and Brain Sciences 45:e183, DOI 10.1017/S0140525X21002351, draw the distinction this chapter is organized around:

• A Pearl blanket is a formal statistical construct — variables in a graph you drew, relative to a model, doing conditional-independence bookkeeping. Cheap, well-defined, unobjectionable.
• A Friston blanket is a metaphysical boundary of a thing — a claim that some real system out there genuinely has an inside, a surface, and an outside.

The field's main error is sliding from the first to the second without paying for it. You define a blanket (free — a modeling choice), then quietly conclude you have discovered where the organism ends (expensive — an empirical claim requiring I(μ;η|b) ≈ 0 in the actual system, which nobody estimated). The target article, its 30+ commentaries and the authors' reply ("The Emperor Is Naked", BBS 45:e219) are worth reading whole; the dispute is live, and this chapter takes no side beyond insisting the slide be named.

And the formal conditions are worse than assumed. Aguilera, Millidge, Tschantz & Buckley (2022), "How particular is the physics of the free energy principle?", Physics of Life Reviews (arXiv:2105.11203; PMC8902446), analytically solved a family of linear Langevin / Ornstein–Uhlenbeck systems and found the FEP's three requirements — a perception–action partition, a Markov blanket, and decoupled solenoidal flows — are in principle independent conditions co-occurring only in a "very narrow space of parameters." Removing solenoidal couplings "precludes ... asymmetric agent-environment interactions, which may be crucial for many living processes"; blanket conditions emerge only "for very particular perception-action interfaces, forcing symmetries in agent–environment interactions that are not expected in living beings."

Read that carefully. In the one setting where the question was solved exactly, the conditions making a Markov blanket exist held only in a narrow, symmetric corner — and the systems this framework most wants to describe (living things: asymmetric, non-equilibrium, solenoidal) sit largely outside it.

Therefore, plainly: a Markov blanket is a MODELING CHOICE you must justify per system, not a fact you may assume. Any use of the partition in this corpus — including M12 — is a typed engineering choice warranted by being useful and explicit, never by having been discovered in nature.

---

Nature as authority here — and the counterweight

The doctrine holds in exactly one form: nature is the authority because it has already run the experiment — a long parallel search under real physical constraints in which the failures were deleted. Convergent evolution is evidence of a constraint-optimum.

The counterweight is mandatory: Gould & Lewontin (1979), "The spandrels of San Marco and the Panglossian paradigm: a critique of the adaptationist programme", Proc R Soc Lond B 205(1161):581–598. Not every trait is an adaptation; phylogenetic inertia, drift, developmental constraint, pleiotropy and contingency all produce features that are not optimal solutions to anything. "Nature does it this way" is a hypothesis generator, never a proof.

That applies to this chapter's own subject, and it is why the fence reads MODELED. The observation that organisms have surfaces is not evidence that the surface is a Markov blanket. Cells have membranes because of lipid physics and history; whether a membrane satisfies I(μ;η|b) = 0 is a different question, and the answer is NOT-MEASURED.

The calibration template — separating the earned from the unearned:

• EARNED: the golden angle, ≈137.5°, in phyllotaxis. Douady & Couder (1992), "Phyllotaxis as a physical self-organized growth process", Phys Rev Lett 68(13):2098–2101, reproduced Fibonacci phyllotactic order in a physical experiment — ferrofluid droplets released periodically into a magnetized dish, repelling and advecting outward — and in simulation. A real number, a real mechanism, a real falsifier. No mysticism required, and none used.
• UNEARNED: "the golden ratio is a universal design law of nature." INADMISSIBLE as stated — unfalsifiable and cherry-picked; it names no observation that would refute it. Recorded with the receipt, not mocked. Someone noticing the spiral is noticing something real; the doctrine attached to it is what fails.

Now apply the same test to this chapter's own claim. "Life has Markov blankets": is it the Douady–Couder kind (mechanism reproduced, falsifiable) or the golden-ratio-universalism kind (pattern asserted, mechanism assumed, falsifier absent)? On the evidence above — Aguilera's narrow corner, and I(μ;η|b) essentially never estimated for a real organism — it is currently closer to the second than its proponents write as though it were. That sentence is the chapter.

---

The numbers (the ratio/frequency table)

Symbol	Value	Units	Scope	Class	Source	Falsifier
μ ⊥ η | b	p(μ,η|b) = p(μ|b)p(η|b)	— (definition)	any system admitting the 4-way partition	MODELED (definitional; assumption = the partition is given, not discovered)	Pearl (1988), Probabilistic Reasoning in Intelligent Systems	Not falsifiable as a definition. The application to system X is falsified by I(μ;η|b) > 0 in X.
I(μ;η|b)	0 (required)	nats	the operable blanket test, any system	NOT-MEASURED for essentially all real biological systems	criterion: Pearl (1988); attempts: Friston et al. (2021) Netw Neurosci 5(1):211–251; Beck & Ramstead (2025) arXiv:2502.21217	Estimate it for a named organism's real boundary. Any I > 0 beyond estimator noise refutes that blanket.
H_μη, H_ημ	0 (required)	precision units	linear Gaussian / Ornstein–Uhlenbeck systems	MODELED — holds only in a "very narrow space of parameters" (assumptions: linearity, weak coupling C² small, homogeneous noise Γ = ς²I)	Aguilera, Millidge, Tschantz & Buckley (2022), Physics of Life Reviews, arXiv:2105.11203	Exhibit a broad, non-symmetric parameter region of a non-equilibrium system where the blocks vanish.
N_rod	92 × 10⁶ (range 77.9–107.3 × 10⁶)	cells	human retina; 8 wholemounts, 7 donors, ages 27–44	OBSERVED-REPLICATED	Curcio, Sloan, Kalina & Hendrickson (1990), J Comp Neurol 292:497–523, DOI 10.1002/cne.902920402	Recount in a comparable cohort; a mean outside the stated range refutes.
N_cone	4.6 × 10⁶ (range 4.08–5.29 × 10⁶)	cells	as above	OBSERVED-REPLICATED	Curcio et al. (1990)	as above
D_cone,fovea	199,000 (range 100,000–324,000)	cones/mm²	human foveal peak; same cohort	OBSERVED-REPLICATED	Curcio et al. (1990)	as above
N_optic	1,159,000 ± 196,000 (range 816,000–1,502,000)	axons	human optic nerve; 22 nerves, 19 subjects, ages 20–75	OBSERVED-REPLICATED	Jonas, Müller-Bergh, Schlötzer-Schrehardt & Naumann (1990), Invest Ophthalmol Vis Sci 31(4):736–744	Recount; a mean outside the stated range refutes.
r_retina	≈ 83 : 1	dimensionless	human visual blanket, order-of-magnitude only	MODELED — assumptions: (N_rod + N_cone)/N_optic across different cohorts, unpaired, no per-eye matching, ignores non-uniform convergence (foveal ≈ 1:1 vs peripheral ≫ 100:1)	arithmetic on Curcio et al. (1990) + Jonas et al. (1990)	Measure both counts in the same eyes. A paired ratio outside ~50–150:1 refutes this estimate.
f_aff (rat)	~80% afferent / 20% efferent	% of fibers	rat, abdominal vagus	OBSERVED-CONTESTED — the widely-quoted "80% of the vagus is afferent," routinely cited outside this scope	Prechtl & Powley (1990), Anat Embryol 181:101–115, DOI 10.1007/BF00198950	Measure the human cervical vagus and obtain 80% ± small. Kronsteiner et al. (2024) did, and did not.
f_aff (human)	sensory 73.9 ± 7.5% (R), 72.4 ± 5.6% (L); parasympathetic 13.2 ± 1.8% / 13.3 ± 3.0%; sympathetic 13 ± 5.9% / 14.3 ± 4.0%	% of fibers	human, cervical vagus; 8 cadavers, immunofluorescence	OBSERVED-CONTESTED — carry with the row above; both positions stand	Kronsteiner et al. (2024), Brain Stimulation 17(3):510–524, DOI 10.1016/j.brs.2024.04.016	Independent replication in a larger cohort; a sensory fraction outside ~65–82% refutes.
N_vagus	~100,000 (light microscopy, 1961) vs 25,489 ± 2,781 (R) / 23,286 ± 3,164 (L) (modern, 2024)	axons	human cervical vagus	OBSERVED-CONTESTED — a ~4× disagreement between methods. The dispute is the finding; the modern claim is that light microscopy cannot resolve unmyelinated fibers	Hoffman & Schnitzlein (1961), Anat Rec 139(3), DOI 10.1002/ar.1091390312; Kronsteiner et al. (2024)	Blinded EM recount across labs on shared specimens. Convergence on either value resolves it.
α_golden	≈ 137.5	degrees	phyllotactic divergence; reproduced in a ferrofluid-droplet physical analogue	OBSERVED-REPLICATED (mechanism earned, not mystical)	Douady & Couder (1992), Phys Rev Lett 68(13):2098–2101, DOI 10.1103/PhysRevLett.68.2098	Run the same repulsion/advection regime and obtain a stably different angle with no parameter change.
n_levels	—	nested blankets	any named organism, established by measurement	NOT-MEASURED	none found	Estimate I(μ;η|b) at each candidate level of one real organism and count the levels that pass.
descending motor signal	proprioceptive prediction, not command	—	vertebrate motor system	HYPOTHESIZED (mechanism proposed, not settled)	Adams, Shipp & Friston (2013), Brain Struct Funct 218:611–643, DOI 10.1007/s00429-012-0475-5	Show descending signals encode forces/commands with no proprioceptive-prediction structure, or reflex arcs that do not discharge predicted state.

---

Falsifier (operable)

This chapter is refuted by any of:

1. The map fails. Exhibit a system in which internal states are demonstrably informed by external states not via sensory states (I(μ;η|b) > 0 with the partition correctly specified).
2. The fence is wrong in the safe direction. Demonstrate that blanket conditions hold robustly and broadly — over wide, asymmetric, non-equilibrium parameter regions with solenoidal flow present — contradicting Aguilera et al. (2022). That would make the blanket more fact than choice, and this chapter's central caution over-stated.
3. The measurement claim is wrong. Produce an existing published estimate of I(μ;η|b) for a real organism's actual boundary. That refutes "essentially never done."
4. The anatomy is wrong. Any table row's value falling outside its stated range on recount in a comparable cohort.
5. The nesting is not a construction. Show a biological hierarchy whose blanket levels were discovered by conditional-independence testing rather than chosen by diagram.

---

Recorded INADMISSIBLE / NEGATIVE (first-class, inline)

• INADMISSIBLE — "a Markov blanket implies a self / awareness / sentience." Unfalsifiable as stated: it names no observation that would refute it. Receipt: the definition quantifies over conditional independence in a partition and mentions no experiencer. The same formalism admits partitions for a Newton's cradle, a burning fuse and the Lorenz attractor — Beck & Ramstead (2025) ran blanket detection on exactly those. If a blanket implied a self, it would imply one for the burning fuse. The claim is the Pearl→Friston slide with a further leap on top (Bruineberg et al., 2022). Recorded, never asserted.
• NEGATIVE — the universal-blanket claim, in the one place it was solved exactly. "Any ergodic system with a Markov blanket…" is frequently read as "…and everything has one." Aguilera et al. (2022) solved linear Langevin/OU systems analytically and found blanket conditions plus solenoidal decoupling only in a "very narrow space of parameters," requiring symmetric interaction loops "not expected in living beings." A first-class recorded negative, not softened here. It does not refute the FEP; it refutes the assumption of universality.
• INADMISSIBLE — "the golden ratio is a universal design law of nature." Unfalsifiable / cherry-picked as stated. Carried with its earned counterpart: the ~137.5° phyllotactic angle is real and has a reproduced physical mechanism (Douady & Couder, 1992). Honor the measured; fence the unmeasured; never mock the asker.
• NOT-MEASURED (the honest empty face), printed rather than filled: I(μ;η|b) for any named organism; the number of nested blanket levels in any real organism; the paired photoreceptor:axon ratio in the same eyes.
• Scope violation, recorded: "80% of the vagus is afferent" is a rat abdominal measurement (Prechtl & Powley, 1990) in near-universal circulation as a fact about the human vagus. The human cervical number is ~73–74%, with a further ~13% sympathetic (Kronsteiner et al., 2024). A number quoted outside its scope is a defect even when the number is right.

---

HONEST FENCE — MODELED

The central object of this chapter — the partition {μ, s, a, η} and the conditional independence μ ⊥ η | b — is a model, and the model's assumptions are the fence: that the partition exists, that it is correctly drawn, that the flow blocks vanish, and (in the tractable Gaussian case) that dynamics are linear with weak coupling and homogeneous noise. Aguilera et al. (2022) show those assumptions bind hard and hold narrowly. Rows inside the chapter carry their own classes — OBSERVED-REPLICATED for the retinal and optic-nerve counts, OBSERVED-CONTESTED for the vagal composition and count, MODELED for the derived ratio, HYPOTHESIZED for predictions-not-commands, NOT-MEASURED where nature has not been asked. A blanket is a modeling choice you justify per system, not a fact you may assume.

---

Not claimed

• Not claimed: that a Markov blanket implies a self, an experiencer, awareness, sentience, consciousness, or a point of view. Explicitly disclaimed. Nothing in the partition names one, and no falsifier is offered because none exists — disclaimed, not tested.
• Not claimed: that the free-energy principle is true, that blankets are universal, or that living systems generally satisfy the blanket conditions. The best available exact analysis says those conditions are narrow (Aguilera et al., 2022).
• Not claimed: that UNI has a Markov blanket in the Friston sense. M12 is a typed engineering partition, status: method, warranted by explicitness and usefulness — never by discovery. No I(μ;η|b) has been estimated for any UNI boundary.
• Not claimed: that anything here raises any UNI rung. A nature citation is never a UNI gate. Every source is published biology and physics; M12's status is unchanged; the honest program position stands at ~2 of 11+ developmental rungs earned.
• Not claimed: that the nesting ladder has a known depth, that the retinal ratio is a measured quantity, or that the vagal afferent fraction is settled.
• Not a target: "full human" and "the next evolution beyond human" are QUAESTIO-APERTA — permanent open questions, never a milestone, never a deliverable. Nothing in the blanket formalism moves them one step closer.

---

Cross-refs: M12 (cookbook/01-kitchen-rules.md, cookbook/02-the-pantry.md) — the typed partition as a method rule. NA-10 — the full scale ladder, organelle → ecosystem, with its measured constants. encyclopedia/NATURE-LEDGER.md — the sovereign class for every row above. encyclopedia/CLAIM-LEDGER.md — the sovereign fence for UNI's own status, which this chapter does not touch.