Note of 2026-10-27: narrowed and promoted
The owner’s ruling on the first status review (ADR-057) carried out the promotion audit’s decision (record/promotion-audit/decisions/THEORY-073.yaml). Every clause of the narrowed title is shown below. The soup’s three choices are documented from the published code, the three placements are in LIT-598 §3.2, and the last clause follows from the formalism, since a fixed graph’s conditional independences carry no variable for how the boundary came about. The unrestricted form, which would reach the later formulations, is THEORY-1097, Proposed, with this entry’s former promote_when. Both dependents survive the narrowing: CLAIM-152 uses the placement clause and CLAIM-194 the made-or-imposed clause, and neither leans on the later formulations.
Source
- Bruineberg, Dołęga, Dewhurst & Baltieri (2020 preprint), LIT-603, read in NOTE-447: §4.2 (the soup re-run from Friston’s code), §5.1 (Friston blankets on an arbitrary graph; co-parents), §6 (the dilemma).
- Raja, Valluri, Baggs, Chemero & Anderson (2021 preprint), LIT-598, read in NOTE-450: §3.1 (every thing, not everything), §3.2 (placement), §3.3 (constitutive self-organization), footnote 17.
- Friston (2013), LIT-526, read in NOTE-421: the account both critiques correct, and the simulation they examine.
The claim
THEORY-066 shows that the blanket formalism takes the internal set as an input. This account says where, in practice, that input and the graph it is applied to come from. Each comes from a choice by the modeller, made before the blanket is traced. A blanket found that way cannot then serve as evidence of where the system ends. That is the sense in which it presupposes the boundary it is used to find. A last clause says what it misses even when well placed: a boundary that the system itself produces looks, to the formalism, like any other conditional independence.
What was actually shown
The soup’s blanket rests on three choices. Bruineberg et al. re-ran the simulation of LIT-526 from its published code. The adjacency matrix was built from electrochemical coupling only, “while other forms of influence included in the simulation (such as Newtonian forces) are ignored”. The eight most densely coupled nodes were declared internal. Steady state was assumed at the time the run was stopped (§4.2, §6). Calling the resulting ring of particles a membrane that the system has is, for them, “a clear example of the reification fallacy”. NOTE-421 had already flagged the chosen k = 8, and this reading adds the other two choices. That is evidence about one simulation, and it is the simulation LIT-526 offers as its demonstration.
Published placements follow the model, not the system. Raja et al. compare three standard applications (§3.2). In a brain, the blanket is the sensory and motor receptors. In a bacillus, it is the membrane and the cytoskeleton. For two pendulums on a beam, the beam is the blanket, while the sensory and active states are each pendulum’s velocity and position, properties of the pendulums rather than of the beam. “The Markov blanket partition is stipulated as being located exactly wherever happens to be convenient for the purposes of the model.” Bruineberg et al. add that co-parents are classed as sensory, active or ignored depending on the author, and that which cause is most proximal is itself model-relative (§5.1).
The formalism cannot tell a made boundary from a given one. A cell makes its membrane. The pendulums do not make the beam, and the brain does not make its receptors in the same sense. Raja et al. argue that the blanket formalism “has no tools to account for” that difference (§3.3). Of Friston’s 2013 simulations specifically, they say they “assume the existence of all the particles from the beginning and, therefore it cannot be said that their blanket states are the product of the activities of internal states” (footnote 17). That is the basis of the record’s narrow correction of LIT-526’s autopoiesis claim.
Two routes, one conclusion. Bruineberg et al. reach it by the distinction between Pearl blankets on a map and Friston blankets claimed for the territory. A realist use needs metaphysical premises “that may in the end be doing all of the interesting work themselves”. Raja et al. reach it from scope. The blanket is a “trick” that supplies the partition variational inference needs, so anything modelled with one becomes an inferrer by construction (§3.5). Raja et al. cite the Bruineberg preprint, and neither paper tests the other.
What this does not say
- It does not say the free energy principle is false. Raja et al. grant it, “given these commitments, we think FEP is true”, and Bruineberg et al. leave the mathematics alone. The claim concerns what a blanket can be used to conclude.
- It does not say no system has a real boundary. Cells have membranes. The claim is that the blanket formalism does not find them. It gets them from the modeller.
- It does not settle the later formulations. Definitions by sparse coupling at nonequilibrium steady state, and blankets whose states change over time, are known to the record only through the critiques’ summaries. They are what THEORY-1097’s promote_when asks about.
- It does not decide the candle flame. LIT-526 excludes the flame because its molecular interactions do not persist (C7 in NOTE-421), and Raja et al. accept that exclusion and turn it against the principle’s scope. If this account is right, whether a flame “has” a blanket is a fact about a model of a flame, and LIT-526 supplies none (NOTE-447).
- It is about individuation, not about the four unities of ADR-024. The question is what makes something one system. No claim is made about selves, persons or the unity of experience.
Connections
- Extends THEORY-066, the formal core: from “the formalism takes the internal set as input” to “in practice the modeller supplies it, and the graph too”.
- Friston’s criterion of the living (LIT-526). The record holds no THEORY stating that a Markov blanket marks a living or lifelike system, so no relation is declared. Both critiques correct the LIT, and this account is the record’s statement of why the criterion does not work as one.
- THEORY-072 offers the alternative. Closure of constraints (LIT-582) draws the boundary at the constraints the system itself produces, which is the “constitutive self-organization” Raja et al. say blankets miss. NOTE-468 records that no paper in the record compares the two criteria on one system.
- THEORY-071 (minimal agency). Barandiaran et al. (LIT-566) require that an agent’s individuality be defined by the system, not by an observer, and they reject statistical measures of influence as a criterion of interactional asymmetry. If this account is right, a Friston blanket meets neither requirement. It is drawn by the modeller, and the arrow structure Bruineberg et al. describe, borrowing Barandiaran’s term “interactional asymmetry”, is stipulated in the graph. It is not a modulation the system performs.
- THEORY-043 (nesting). LIT-526 says the blankets of animals enclose those of organs and cells. NOTE-421 reads that nesting as bearing on THEORY-043’s anti-nesting question. On this account, which blankets are drawn is chosen, so a nesting of blankets cannot by itself decide that question.
- Coordination dynamics (LIT-601). Raja et al. note that the coupled pendulums are already explained by their relative phase, the HKB treatment read through Kelso 2021. A blanket adds only a Bayesian redescription.