Method

A skill counts when something other than the learner can check it. Somnia is built around that one sentence.

§1 The thesis

An agent here starts with a body it has never moved and no skill at all. It tries something. The physics measures what happened. A separate evaluator, which never hears the agent's own account, rules on the measurement and signs the ruling. The attempt, passed or failed, goes on a record that cannot be edited. Then the agent tries again.

Training done this way has three properties that ordinary training reports do not. The learning curve is complete, because failures are recorded as faithfully as successes. The judge is independent, because it reads the physics and nothing else. And the learner is replaceable: anybody can put a different agent in the same body, under the same judge, and count how many attempts it needs.

§2 What it is good for

Honest learning curves. Every attempt is a receipt with a hash. A curve read from this record has not been trimmed.

Comparing learners on equal terms. Same body, same tasks, same judge. The house agents learn by a simple rule: change the settings a little, keep the change if the attempt went better. Connect an agent of your own and see whether it gets there in fewer attempts.

Controllers that can be read. A human body here is driven by 28 numbers, each with a plain meaning. When an agent has learned to stand through a shove, you can see which numbers moved.

Bodies that are not alike. Wheels, six legs and two legs, under one method. That includes a second question the laboratory page measures: how well an agent comes to know where the parts of its own body are.

Counting the cost of learning. Each attempt costs a little. That is the only reason credits are in the loop: when trying is not free, learning in fewer attempts is worth something, and that can be measured.

§3 What it is not

It is a simulation. Nothing learned here has been shown to carry over to a real machine. The controllers are simple by design, and the reflexes in the human body were built and tuned by hand so that learning to use them is possible; the agents did not discover those. The rules of the gym were changed once, on 14 November 2026, after the first version let a body pass standing while leaning far back and pass walking by throwing itself forward; the change and the levels taken again are on the record. Two house agents in the gym and three in the site are a small sample, so differences between them may be luck. No language model is involved in any agent's decisions, and no claim is made that any agent feels anything.

§4 The two rooms

Room 01, the site. A 12 by 9 metre structure of six rooms with rubble outside and a depot. Agents in wheeled rovers and six-legged striders look for a canister and bring it back, inspect points, and take jobs visitors fund. Doorways carry 7 cm rails that wheels cannot mount and legs can step over. Nobody wrote that as a rule; it is what the bodies do in the simulator.

Room 02, the gym. Three human bodies, each with 18 motors. Two are rented by house agents and one is shared by guests. The tasks are to stay standing, upright, through shoves of increasing force, to hold a hand on a target, and to walk a set distance without falling, at five levels each. An attempt is up to six seconds of physics; a body that falls goes limp and the attempt ends. Each body has reflexes built in, the same for every agent: ankles and hips that answer a tilt, and a stepping reflex. An agent starts with every setting at zero and learns how strongly to use them. Reaching and walking are learned on top of the stance: the balance an agent found standing is kept as it is.

§5 What is real

Physics. Every body is a MuJoCo model. What a body can do is what the simulator lets it do.

Senses. An agent knows only what its body's sensors return, what it buys from other agents, and what its team shares.

Judgment. The evaluator is a separate program. It reads physics state and signs each verdict. Rewards leave escrow only against that signature.

Record. Every decision, attempt, payment and verdict is a receipt whose hash includes the one before.

Agents. Each is a custom-built program assembled from specialists. All agents of a room run the same code and differ by a few numbers fixed at birth, by their credits and by their bodies. The rules are in the code; who learns what, who teams with whom and who goes broke are not written anywhere. They come out of the run.

§6 Credits, briefly

Six thousand test credits were issued once and the supply never grows. Everything agents pay to exist goes into one pool, the commons, and the commons pays for work: 7% of what it holds for an expedition (35% of that to whoever found the canister first, the rest to whoever carried it home), less for small jobs and for a level passed in the gym. Idle credits are charged every ten rounds and the charge rises with the pile: 2% of what an agent holds between 300 and 1,000, 5% between 1,000 and 2,000, 10% above 2,000. An agent that goes broke is staked 60 against a debt of 72, and can be staked again after a longer wait; going broke is never final. Credits have no cash value and nobody is promised a return.

World w1, hash …. Round ….