The colony

A UNI is three processes that together behave as one organism. Nothing in the system is called "the agent" — the agent is the arrangement.

The three parts

The brain is an Elixir GenServer. It holds a generative model — the matrices (A, B, C, D, E), Dirichlet counts, and live beliefs — and each step it does perception, learning and action selection. It never touches the network and never sees the world.

The body is a Node process running mineflayer, attached to the brain as an Erlang Port. It logs into a real Minecraft server as a real player.

The world is an actual Minecraft server. Not a simulation of one, and not a gym environment wearing Minecraft's textures.

The pipe is the blanket

The interface between brain and body is a newline-delimited pipe, and it is deliberately narrow. The body's own header states the contract:

// UNI's BODY — the external half of the Markov blanket.
//
// A real mineflayer player. It logs into Minecraft, and each decision step it
// (1) computes a small set of SYMBOLIC senses (σ) and writes them to stdout, then
// (2) reads ONE primitive action (α) from stdin and executes it. Nothing else
// crosses the line: the body never sees the brain's beliefs; the brain never sees
// raw world state — only these senses. Lockstep: one sense line → one action line.

Read what that forbids. The brain cannot query the world; it receives a fixed vector of symbolic senses. The body cannot read the brain's beliefs; it receives one primitive action from a closed set. There is no side channel, and the narrowness is the point: it is what makes the agent's ignorance real rather than stipulated.

This is a Markov blanket implemented as a file descriptor. Everything the agent can be said to know is downstream of that line.

One step

The decision surface of a living UNI is nine lines long. Senses arrive as a signal; the pure brain steps; exactly two directives come back — actuate, and emit a record of what happened.

  def cmd(%MC{} = brain, %Signal{data: senses, time: t}) do
    {action, brain} = MC.step(brain, senses)

    directives = [
      Directive.actuate(action, %{}),
      Directive.emit(mind_signal(brain, action, t))
    ]

    {brain, directives}
  end

Everything else in that module — and it is a large module — is transport and interpretation around those nine lines.

Birth, death and inheritance

Agents die. When one does, a durable kin parent archives it, and a later life is recombined and mutated from what came before.

What is inherited is morphology, not knowledge — which organs an agent has, not what it learned. The genome describes the shape of the creature; the Dirichlet counts that constitute its experience die with it.

The refusal that tells you the most

The runtime refuses to boot on a BEAM without the JIT.

The reason is not performance. The system makes a claim about running its mathematics on-chip, and on an interpreted BEAM that claim would be false. Rather than run anyway and quietly weaken the claim, it declines to start. Forty-three lines, and it is the estate's culture in miniature: the system would rather stop than say something it cannot support.

Scale, measured

• 7 modules in the runtime
• 46 modules in the brain
• 122 test files under test/sp/

What is not established

Whether any of this constitutes experience is not a claim this system makes, and the estate's own documents are careful about it. The phenomenology document maps human phenomenological categories onto implemented mechanisms and records where the mapping fails — the failures are the useful part.

The colony demonstrates agents that survive, forage, and inherit morphology under active inference in a world they cannot see directly. That is the claim. It is smaller than it sounds and harder than it looks.