A Markov blanket is a thing's skin, written as a rule: whatever the inside learns about the world comes in through the skin, and whatever it does to the world goes out through the skin. So once you know exactly what the skin is doing, the inside has nothing more to tell you about the outside.
Drag in the tank to stir. Drag outside it to move the camera.
The picture of the ink drop is Karl Friston's: ink in water disperses, because there are vastly more ways for it to be spread out than gathered, and anything that lasts has to keep doing the opposite. Everything on screen is computed as you watch. This is a cartoon of the mathematics, not a model of a cell.
A flat slice of the tank, 240 by 150 cells on a staggered grid (each cell about 1.3 mm of a 32 cm tank), solved on the graphics card. The scheme is stable fluids (J. Stam, SIGGRAPH 1999): the velocity is carried along itself by tracing back along the flow, a small viscosity and a faint drag are applied, and a pressure solve (40 Jacobi sweeps, started from the last answer; once the film is undercranked and the flow is slow, 8 sweeps and a doubled time step) makes the flow incompressible. Ink is carried by an upwind scheme in flux form (with van Leer's limited slopes, to keep the plume crisp) and diffuses; a scheme in that form cannot create or destroy ink, so the "ink in the tank" figure should stay at 100% but for rounding. Ink is a little denser than water. Five slow gyres, each waxing and waning on its own clock, stand in for the convection that every real tank has. Entropy is the Shannon entropy of where the ink is, taken over the 36,000 cells; its ceiling is log2 of 36,000, about 15.1 bits, reached when every cell holds the same share. Once the plume has passed the film is undercranked as far as your machine allows, and the slate says by how much.
Sixteen cells in a ring: eight sensory (gold), eight active (red). A small heater hangs in the tank and wanders slowly; warmth spreads from it far faster than ink does and leaks away through the glass. Sensory cells each read the logarithm of the warmth where they sit, with a little noise. The inside holds a Gaussian belief q about where the heater is, measured from itself: a mean and a spread. Its generative model says: the heater is somewhere nearby (prior: a Gaussian centred on itself, 60 cells wide), and a sensor at distance r from it should read b − x − ½ ln(x + 0.1), with x = r/28, give or take 0.25 (roughly how warmth thins out from a point in a thin tank that is losing heat). The belief moves down the gradient of the variational free energy F = complexity + inaccuracy, where complexity is the divergence of q from the prior and inaccuracy is the expected squared misfit of the eight readings (between them they give the level of the warmth and its slope), weighted by their assumed precision (less a constant that depends only on units, and with the curved part of the model left out). Active cells are set by the belief and nothing else: each pumps stray ink back inward at a fixed internal setting, and hauls the creature's ink through the water in proportion to how far the belief says the heater lies in its direction, pushing a little water back as it goes. The currents carry the creature as they carry everything else. The gold and red you see are pigment in the water, held in place by the same pumps, so when the actuators are cut the skin dissolves with the rest.
The three switches cut real lines of the program. With "sensors tell the inside" cut, the readings are still taken but never reach the belief, which sinks back to its prior. With "inside drives the actuators" cut, the active cells receive only their own noise. With "actuators work on the water" cut, they fire and nothing happens.
Every sixth step the page records the belief (inside), the true offset of the heater (outside), and the 24 numbers of the skin (8 readings, 8 pump settings, 8 pushes). The first figure is the ordinary correlation between inside and outside over the last 360 records. The second is the partial correlation: both are first predicted as well as a straight-line fit allows from the skin's numbers alone, and what is left over is correlated. Because the active cells carry the belief out so directly, many designs would pass this test. It shows what the definition says; it does not show that anything is alive.
The third dimension (in a real tank the drop makes a vortex ring; in a slice it makes a vortex pair). Surface waves, the glass's grip on the water, and any chemistry. A real cell's blanket is not a tidy band of two kinds of cell, and here the creature's shape and its pumps are given by decree rather than earned. Warm water is not allowed to rise, as it would. Its model of the warmth is deliberately rough, so its guess of the distance is biased. The free energy principle is a principle and a way of modelling, not a finding that one experiment could refute; its author says so. How far the Markov blankets drawn in models correspond to real boundaries is disputed, and conditional independence given the present state of a blanket does not hold for every system that has one drawn round it.
K. Friston, "Life as we know it", J. R. Soc. Interface 10 (2013); K. Friston, Nature Reviews Neuroscience 11 (2010); T. Parr, G. Pezzulo and K. Friston, Active Inference (MIT Press, 2022); M. Kirchhoff and colleagues, J. R. Soc. Interface 15 (2018); J. Bruineberg and colleagues, "The Emperor's New Markov Blankets", Behavioral and Brain Sciences 45 (2022); M. Aguilera and colleagues, Physics of Life Reviews 40 (2022); J. Stam, "Stable Fluids", SIGGRAPH 1999.