The project did not start with a language model at its centre. It started with the opposite conviction.
The bet we made first
Path A was an attempt to step outside the language model entirely. The core unit was meant to be a persistent, self-regulating dynamical substrate: an active-inference generative model with homeostatic setpoints, decay, and the kind of internal tension that, in the cybernetics tradition, is supposed to be the cognition. The language model was deliberately demoted to a narrator: a mouth, not a mind. It could read out the substrate’s state in plain language, and it could not touch it.
The appeal is easy to state. If cognition really emerges from persistent self-regulating dynamics rather than from a larger model, then the substrate is the organism and the language is just a window onto it. We built that window, wired up the substrate, and let it run.
What actually happened
It converged into a deterministic attractor. Run after run, the substrate settled into the same fixed repeating state and stayed there: the dynamics that were supposed to generate open-ended behaviour instead drained into a single basin and went quiet. The narrator dutifully reported the same thing forever. There was nothing for it to narrate.
This is, to be fair, the failure mode the field has warned about for decades: a self-modifying dynamical system collapses, drifts, or loops, and the hardest problem is never intelligence but stability. We hit the collapse case directly. The substrate had no requisite variety to grow into, and the inference loop found its silent equilibrium.
The hardest problem was never intelligence. It was stability.
The decision
We did not try to rescue it that week. On 2026-05-20 we deprecated Path A.
The honest reading is that Path A failed on its own terms: the substrate-as-agent design produced a deterministic attractor, not a life.
Not: this does not prove the dynamical-systems route to cognition is wrong in general. It proves our first substrate, at the scale and tuning we could run, fell into a fixed point. Others have built richer substrates; we ran out of variety before we ran out of ideas.
The pivot was to invert the architecture. Instead of a language model narrating a substrate that does the thinking, we would treat the language model as the organism and the substrate as its body, world, and memory: the thing it must survive in, not the thing that thinks for it. The narrator becomes the cognition; the substrate becomes the environment that can injure it and the biography that makes it itself.
That is Path B, and everything since this entry is built on it. The old active-inference machinery still runs as instrumentation. It is no longer the agent.
Why this entry exists
A clean notebook would skip the architecture that died on day one. We are keeping it because the pivot is the most informative thing that happened that week. We made a strong, specific bet, the world refuted it inside a single working build, and we changed direction rather than tuning the collapse away. If the project is worth following, it is because of decisions like this one. And there will be more of them.