Most of what is called artificial intelligence today is built to carry out a human intention. That is legitimate, and it is not what happens here. The question this work stands under is older and narrower: what conditions would let something develop a self at all — from below, out of a minimal core, the way a body grows out of cells and not out of a drawing of a body.

One sentence carries both strands. Cognition costs. Where the outside circumstances do not demand it, it is avoided. In people that is an observation. In a primordial soup of executable code it is a measurement: the blind route wins reliably, as soon as the world allows it. Both are the same sentence — and it is the reason this is about conditions and not about algorithms.

The commitment to ordinary hardware follows from the same place. Not because scale is wrong, but because an intelligence that exists only in somebody else's data centre is not available to the person it is supposed to stand beside.

Can perception, coupled to action, arise on its own under pure replication selection? With no score, no fitness, no reward for seeing. Almost every project of this kind smuggles in an objective function and thereby builds the answer into the question.

What is selected for must not be measured.
What is measured must not be selected for.

A regression test twists every observed counter into absurdity and checks that the world's trajectory does not shift by a single round. All measuring is bookkeeping, and it never reaches the physics. An organism lives as long as nothing comes for it, and computes as long as its energy account holds out. There is no more selection than that.

What was expected was a statement about worlds — under this one cognition appears, under that one it does not. What came out was more useful: eight things a substrate has to bring with it before the question becomes measurable at all. Every one of them was an assumption built in as an implementation detail, readable in our own code — and noticed as a wall only once somebody walked into it.

And one organism managed it. f001 arose over 1,205 generations from an ancestor with no threshold, and places its children with a quality of 0.877 where chance sits at 0.50 and the best hand-built design at 0.486. The mechanism predicts the number; switch the sensor off and it falls to 0.321.

To do it, it uses none of the tools that were built for it. The comparison the instruction set never gave it, it found in the error path of memory allocation: a trial request that fails exactly when the place is too poor or already taken. One test, two conditions. The organism turned a hand into an eye.

The limits are as large as the finding: one organism, one seed, one physics. f001 senses the target it steers towards — in this world the only possible form. And the north star, complexity that does not saturate, has not been reached. Mean genome length stayed flat across the whole arc.

Not a tool that is called and answers. The computer is the body: its load, its memory, its heat, its sleeping and waking are the world being perceived — and to act on that world is to act on itself. When it thinks, it grows warm, and its own sensors measure it. This is deliberately not corrected away.

The claim lies on a different axis from that of language models: adaptation rather than knowledge. A language model knows more about the world than this system ever will. This system knows more about this one computer than a language model ever will — and it knows it better tomorrow than today.

The core carries clear, persisting states rather than sequences of tokens. Language and retrieved knowledge are allowed at the edge, where they are useful and optional; they are never the identity and never what decides.

What has been measured so far. First the agent had to be able to learn at all: seven conditions had to come together, and every single fix did nothing while another was still missing — the last was a number in a configuration file. Then a condition on its own process rather than on the world — something has to happen that it can see — gave its drive direction: four independent runs agree at 0.81, where curiosity and empowerment alone stay at 0.26.

Since then it lives in the real machine. It reads its file system, writes with a hand of its own, and sets its own learning effort and its own clock. Sixteen times I gave it more — more world, more senses, a teacher, a key that turns sixteen bytes into a megabyte. It used all of it, and none of it moved its drives. It formed a preference only once something was expensive: a price on its own compute shifts its policy by 4.36 spans onto exactly that quantity, in both seeds, repeated in two worlds.

The gap that stays open: nothing in its world happens by itself. It consists of noise it cannot move — something changes on every step, and not one of 380 channels is more than 11.5 per cent predictable — and of its own position, which it knows because it set it. What is missing is a process that happens by itself and follows a rule it could learn. That is the next stop in the project.