The roadmap
The apparatus exists and runs. From here the work is to make the substrate richer, turn it into an instrument that yields defensible numbers, and only then attempt the hard thing. Software first; the body, last. Stated in order, with the stages that could stall it.
Each phase is gated on the one before it surviving its controls. We do not put dates on research we have not done, and we do not commit to the next gate until the data from the current one is in. The order is fixed; the timeline is not.
The instrument is real, and the method that follows is laid out below.
Not: a roadmap is not a result, and an order of intentions is not a promise. Each phase is conditional on the prior one holding under a hostile control. Any of them can stall; the notebook will say where.
A continuous, mortal body lives or dies on its own physics inside a bounded world: it has integrity and energy, leaves marks and songs, and carries its biography across a life and past its death. A frozen language model is wired into that body as a bounded brain, reading a compact summary of its state and proposing only small parameter nudges (a lean toward repair, or toward seeking food). It cannot move, eat, or act directly. The runtime exists, runs on a desk, and persists to disk.
A frozen language model read at each step as a bounded brain: it sees a compact summary of the body's state and proposes a small parameter nudge toward repair or toward food. It does not pick actions, move, or eat; the body does. The brain lives inside the organism, not the reverse.
A bounded grid world the organism cannot see out of: food, hazards, water it can fish from. Hazards take integrity; food restores energy; the world updates as it acts.
Persisted to disk and surviving restarts. Each life writes one paragraph on death; songs persist at the cells where they were sung; faint traces decay over time.
When a body dies, what it changed in the world remains for the next one. Successors inherit what came before (and absorb, invert, or reinterpret it) through clearly separated voice and lesson channels.
● Status: apparatus built and running · first controlled experiments in flight · zero life-properties demonstrated
Two threads run in parallel. The first adds primitives that change what kind of being the organism can be: not features bolted on, but new ways for it to remember, transmit, and inhabit. The second turns the apparatus into a measurement machine: mechanism on versus off, on matched seeds, against shuffled controls, producing numbers rather than anecdotes.
An offline rest-and-replay register; deliberate bequests addressed to whoever comes next; named places that turn terrain into authored geography; named elders that turn a list of lives into an oral tradition; permitted silence. Each changes the organism, not just its action set.
A formal ablation matrix over what already exists. For each mechanism: off, on, shuffled labels, stale labels, prose-only, different model, different substrate. The deliverable is an effect size with a confidence interval, against a control that did not move.
Viability, homeostasis, body schema, place & autobiographical continuity, lineage transmission, each mapped to existing, cited math and a first-pass operational definition extracted from the logs.
If place and body coupling hold, a richer sensory channel: a camera and audio read as substrate-classified labels, on the same architectural principle. A deeper world to be coupled to, still in software.
The number the instrument is built to produce asks how much of the organism's behaviour the substrate explains, holding the model fixed.
Not: the borrowed math is not our contribution, and a metric is not a discovery. We cite viability theory, active inference, and the rest; what is ours is the operational synthesis and the controls. A strong number still has to survive its shuffled twin.
The hard turn. The lived loop produces grounded tuples: what was sensed, thought, done, and what followed. Those become bounded, sleep-time weight updates, with matched controls and a drift budget. The same five axes that measure the organism then measure the adapted model. The question is deliberately not "can we make it more alive?"
The measurement runs that produce the science also produce the training data; the same extractors that score the organism score the adapter. No new conceptual machinery: the loop closes on itself.
Does adaptation improve substrate coupling under held-out perturbations, or does it merely train the model to imitate organism-flavoured prose? A sharper test than any narrative quote could pass.
Real substrate, absent substrate, shuffled substrate. If organism-prose stays high when the substrate is empty or its labels are randomised, the model learned a roleplay distribution, not coupling. Both failure modes are publishable.
If adaptation improves coupling and not just imitation, that is a real architectural finding about frozen-model systems.
Not: this is not self-improving AI, not biological learning, not a claim the organism gets "smarter." It is a bounded adapter measured against controls. A negative result (learned roleplay, no coupling) is a planned outcome, and the work proceeds under either.
The phases above are the means. These are the properties that decide whether the larger bet, a synthetic organism, lands at all. Each is a genuine open question, and each comes paired with the claim we explicitly do not make about it.
Behaviour that works to preserve the organism's own continuity against decay. The test is causal: remove the capacity and life measurably shortens.
Goals that arise from the organism's own dynamics and its mortality: behaviour that protects continuity, and adaptation to change it was never told about, rather than hand-coded if-then rules.
Heritable variation across a lineage under mortality-driven selection. We have observed inherited registers absorb, invert, and reinterpret across generations; whether that becomes open-ended is unproven.
Sustaining itself over long horizons in a body, ultimately a physical one. It comes last in the sequence, after the software organism already works, as the final transfer test under contingency it was never tuned for.
If every property holds (measured, controlled, in public), that is the strongest operational case we know how to make.
Not: we will never claim the system is alive, conscious, or that it experiences anything. Boden and Pattee hold that software running on hardware it does not itself produce may never be alive in the strong sense. We keep that ceiling and aim only at the operational target. Autopoiesis in its full sense is explicitly out of scope.
Built and running; the first controlled experiments in flight. No life-property has yet survived a hostile control: the honest starting line.
Primitives of being, the ablation matrix, the five axes, and (if coupling holds) a software sensorium. Numbers, not anecdotes.
Bounded adaptation from the organism's own life, measured on the same axes, against the absent-and-shuffled control triplet.
The working software organism transferred into a physical robot, only after it already holds together in software. Never first.
This is a long-horizon research bet, and it might not land.
Not: we are not promising a living robot on a schedule, and the robot is deliberately the terminal step, not a parallel track. Companion or chatbot framing is the wrong attractor and is out of scope. Every stage is published with its controls and its failures; if it stops, the notebook will say where it stopped.
If we get there, you will have watched it happen here, one controlled result at a time.