01
Body
A continuous, mortal system: energy, accumulated damage, perturbation, and death. It runs on its own physics and can die with no brain at all.
The apparatus
A body, and a model demoted to its brain.
We did not train a model, and we did not change one. We built a body the model lives inside. The body is a continuous, mortal system that runs on its own physics; the language model is demoted to a bounded brain, an organ that can only nudge the body's parameters. The whole project then turns on one falsifiable question: does that brain actually do anything?
The model is an organ, not the organism.
Not: this is a course-correction. We spent a long time trying to make the model itself be the organism, but a transformer is inert between calls and reads text rather than a world, so that was a category error. The notebook records why we stopped.
The parts
What the apparatus is.
02
Brain
A language model used as the brain. In the baseline it is frozen and never fine-tuned; we also test online adapter updates that let its weights slide during a life. It sees a compact summary of the body and proposes small parameter adjustments. It cannot move, eat, or act directly.
03
The interface
The brain's entire influence is one narrow channel: a state summary in, small parameter leans out (toward repair, toward seeking food). Everything it does, it does through that channel.
04
Mortality
Death is real. The body can deplete its energy or accumulate fatal damage and the run ends. The brain's job is to help the body survive, not to keep itself alive.
05
The controls
The honest question is whether the brain matters. Four controls answer it: no brain, an invalid-signal brain, a randomized-weights brain, and a valid-random brain. A result counts only if the trained model beats all four.
06
Discipline
Before any long task we check a hand-coded optimal policy can survive it; a failure on an unsurvivable task tells you nothing. Parser bugs that credit ambiguous output are caught, and the arm is binned.
The dividing line
The body acts. The brain only leans.
The body is where life or death happens, on its own physics. The brain never acts in the world: it cannot move, eat, or repair anything directly. It can only lean the body's parameters a little, and then the body's own dynamics carry it. Holding that line is what makes the question clean. If a trained brain outlives its controls, the credit is the model's learned prior, not our code acting in its place.
The model's only power is a small lean on the body's parameters.
Not: a lean is not agency. The brain has no hands; what survives is the body, and we measure only whether the model's lean does better than noise.