Jelle Bruineberg, Krzysztof Dołęga, Joe Dewhurst and Manuel Baltieri, Behavioral and Brain Sciences 45:e183 (2022; online 22 October 2021); first public as a PhilSci-Archive preprint, December 2020.
The brief’s citation (BBS, 2022) is correct; the authors’ names are as Crossref gives them.
Which version was read. The brief asked for the target article and allowed the PhilSci preprint. The target article is not lawfully readable here beyond its abstract, so this entry rests on the 2020 preprint. The abstracts of the two agree on the Pearl/Friston distinction and on an instrumental-versus-realist dilemma. The 2020 abstract states the dilemma as two research programmes; the BBS abstract frames the argument around the free energy principle’s claim to integrate the life sciences and names the programmes “inference with a model” and “inference within a model”, the terms the 2020 body already uses (§6.2). The authors say the published version cut technical material, which is the material this record’s relation to LIT-526 rests on, so a reader should not assume every detail below survives in the BBS text.
Key takeaways
- Two constructs under one name. A Pearl blanket is the parents, children and co-parents of a node in a Bayesian network; it simplifies mean-field variational inference and says nothing about the world. A Friston blanket partitions a system into internal, sensory, active and external states and is offered as the agent–environment boundary itself.
- The partition is not in the graph. Applied to an arbitrary network, the four-way labelling requires first choosing which node is “internal”; different choices give different blankets, and the sensory states need not coincide with the observed variables. Co-parents are treated inconsistently across the literature, and the patellar-reflex example shows they can stand for external causes indistinguishable from internal ones.
- The primordial soup, re-run. In LIT-526 the blanket is found after building the adjacency matrix from electrochemical coupling only (ignoring the Newtonian forces in the same simulation), assuming steady state when the run is stopped, and fixing the number of internal particles before spectral clustering. The resulting “membrane” is attributed to the system: “a clear example of the reification fallacy”.
- The dilemma. Instrumental blankets (“inference with a model”) are legitimate but cannot settle where minds or organisms end. Realist blankets (“inference within a model”) could, but only with additional metaphysical premises “that may in the end be doing all of the interesting work themselves”.
Standing in the record
Held in both records. Mathematics of Meaning holds this paper as MOM-LIT-023, read there for its own question: that corpus adopts the blanket for the self-boundary and the organism–artifact cut while citing this critique. Its verdict, “a property of a model, not a boundary in the world”, agrees with this reading. The two entries are separate readings (ADR-013, ADR-037).
Filed on 2026-10-03 at the owner’s request, in the free-energy part of the agency batch. No anthology topic holds it. It explains variational inference and mean-field approximations, but as background to a philosophical argument about free-energy theory, and it carries no instruction for machine-learning practice.
Relation to LIT-526: corrects. “Life as we know it” claims that
Markov blankets form in its simulated soup and that the internal states
they enclose appear to infer and to maintain themselves; it uses the blanket
as the mark that separates a lifelike system from its surroundings. This
paper shows the blanket in that simulation is not discovered in the
dynamics but produced by modelling choices the paper does not justify:
which interactions define the graph, that steady state has been reached,
and how many internal states there are. So the claim that the soup
possesses a boundary in the world, rather than one drawn on the modeller’s
map, does not follow from the simulation. That is a correction of a claim
LIT-526 makes, by argument and by reproducing its figure from its code; it
is not an empirical refutation of the lemma, which this paper does not
dispute. NOTE-421 had already recorded the chosen k = 8 and the
redescriptive character of the proof; this paper supplies the general
diagnosis.
Where else it bears:
- Raja et al. (LIT-598) cite this preprint and make a parallel argument from a different side: the blanket is a “trick” that makes any thing a Bayesian inferrer. Neither paper tests the other.
- Barandiaran, Di Paolo & Rohde (LIT-566). This paper describes the arrow structure a Friston blanket assumes (no edge from internal to sensory, none from external to active) as an “interactional asymmetry”, citing them. Their definition of agency demands that the system itself, not the modeller, fix its individuality, which is the point this paper makes about blankets.
- Organisational closure (LIT-582). A boundary criterion whose boundary is produced by the system’s constraints, not chosen on a graph. This paper does not discuss it.
- Sophisticated Inference (LIT-577) is an example of the “inference with a model” programme, an agent’s generative model over a task, and is untouched by the critique as an algorithm.