Origins
This began as research, not code.
Through 2025 the work was a single question: what if the whole scaling, centrally-trained, objective-maximizing paradigm is the wrong road? The bet underneath it: that cognition comes from persistent, self-regulating dynamics, not bigger models.
A question that led back through cybernetics.
The bet led back through cybernetics, second-order systems theory (Maturana and Varela's autopoiesis), and active inference. Not a search for a better transformer, but for the tradition that takes maintaining oneself as the unit of cognition, rather than mapping inputs to outputs.
A mouth, not a mind, and a dead end.
The first attempt put no language model at the centre at all: the model demoted to a translator, "a mouth, not a mind." It ran into a deterministic attractor and was deprecated. The lineage was right; that particular instantiation was not.
Keep the model, as an organ, not the mind.
The reframe that worked kept the model but demoted it: a bounded brain, an organ inside a continuous, mortal body with its own world, energy, and state. The body lives or dies on its own physics; the frozen model only reads a compact summary of that state and proposes small parameter nudges (a lean toward repair, or toward seeking food). It cannot move, eat, or act directly. It is still useful for its one real gift: it can read out what is happening inside, closing the legibility gap that left earlier artificial life unreadable. That became Synthena, built and documented in an intensive sprint in 2026.
The body lives or dies on its own physics; the model is a bounded brain that only nudges it.
Not: the model is not the organism, and Synthena is not "the model with a prompt." The body runs on its own physics; the frozen model reads a compact summary of that state and can only propose small parameter nudges (a lean toward repair, or toward food). It cannot move, eat, or act directly. But that learned prior is load-bearing for survival: it beats four controls across two model families. That is not life, consciousness, or self-maintenance, and zero life-properties are demonstrated.
Two paths, one architecture.
The same primitive kept surfacing: a runtime where the body acts on its own physics and a frozen model is kept out of the action hot path, proposing only bounded nudges rather than driving the system. That is the shared shape the decision-runtime work, Syntra ↗, converges on. The research and the runtime arrived at one architecture from opposite ends.