Primordial-Core · soup · PC1–PC17 · closed July 2026
What a substrate has to bring with it
Can cognition — perception coupled to action — arise on its own under pure replication selection? An arc of seventeen experiments, and the only organism that managed it.
This is a working record, not an essay. It is here because the numbers in the Substrate strand would otherwise be assertions. Anyone who only wants to know what came out should read the first section and the last; everything in between is the road, wrong turns included.
The question
The guiding vision looks for a self that arises from below out of a minimal core — the way a cell arises from a genome — rather than for a serving tool. But that question can only be asked if it has not already been answered.
Almost every project of this kind smuggles in an objective function: reward seeing, reward efficiency. With that, the answer is built into the question. Here there is no score, no fitness, no evaluation. An organism lives as long as the reaper does not come for it; it computes as long as its energy account holds out. Whoever copies more often is present more often. There is no more selection than that.
What is selected for must not be measured.
What is measured must not be selected for.
So that this does not stay a declaration of intent, it is a regression test: it 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.
The machine
Enough to follow everything that comes after. soup stands in the Tierra/Avida line: a metabolic virtual machine in which genomes copy themselves and compete for energy and space.
Arena — a ring of words
A flat ring of instruction words. Memory is scarce and rivalrous — it is the population limit.
Genome = code — 32 instructions, no gaps
Five bits, 0 to 31, without a gap. Every bit flip yields another valid instruction. Mutation cannot produce a syntax error, only a different meaning. That matters later, and as a drawback.
Addressing — by templates, not by addresses
A jump searches for the nearest complementary template pattern. That is why insertions and deletions survive: a jump target travels with the code instead of pointing into nothing.
The ancestor — 49 words, written by hand
It finds itself, measures its own length, requests memory, copies itself word by word into it and divides the child off. That is the last time a human writes a genome.
Metabolism — energy out of the real machine
The energy inflow of the whole soup is modulated by CPU idle time, free memory, real GPU temperature and genuinely measured memory latency. It is distributed in equal shares per living organism — any other distribution would be an objective function smuggled back in.
Selection — out of scarcity, not out of judgement
Energy is a rate limiter, memory the population limit. Whoever has spent their account waits; they do not die of it. But the reaper keeps walking and takes the oldest. Whoever needs many instructions per replication is caught before their child is finished.
The result
What was expected was a world condition: under X cognition emerges, under Y it does not. What came out was something else, and it is more useful.
The hardest limits in this project were never laws of evolution, but assumptions built in as implementation details — noticed as walls only once somebody walked into them.
Every one of those walls was readable in our own code before it was measured. Not one was noted down as a restriction. The result of the arc is therefore a ladder of eight conditions — each rung met singly, measured, and paid for dearly. The overview first; the wrong turns that made them visible come further down.
- 1 An actuator that can read a place. Allocation read the size of the child, never its place. Perception had literally nothing it could have acted upon. PC2 · after which perception wins by 10.8×
- 2 A world without a cheap blind shortcut. Periodicity and spatial correlation are compressibility. Anything compressible can be baked in structurally instead of being asked for through a sense. PC5-B · PC6-A · PC15-B
- 3 A world that demands more than one bit. One good stripe, one bad stripe — one bit carries at most one bit of coupling. That is arithmetic, not measurement. PC6-B · graded quality 0…15
- 4 A path walkable in both topologies. A road can be fitness-monotonic and still impassable: fitness neighbourhood and mutation neighbourhood are two topologies, and only their intersection is a road. PC7-B · three bit flips → one bit flip
- 5 Enough generations — and the unit is not the second. Four million rounds are about a hundred generations. Every cost parameter that shapes the ecology silently rescales evolution's unit of time. PC7-C
- 6 A precursor that survives until it is climbed. A capability that does nothing yet is neutral — and what is neutral is taken apart by drift in about three hundred generations. PC7-D · and the correction we had to make to ourselves
- 7 An action space with more than one verb. An organism could write nowhere except into its own child. The entire action space was one verb: where do I put my child. PC8-A · the runner-up
- 8 A channel with an addressee — and an alphabet for messages. Making perception expensive creates no market for signals. And in living code all 32 instructions occur — no value is free to serve as a marker. PC16 · PC17 · the wall this arc ends at
The positive finding
One case in twenty-four cells. It carries anyway, because it is fully explained.
f001 arose under pure replication selection, over 1,205 generations, from an ancestor whose threshold was zero. It reaches a placement quality of 0.877 — against a chance value of 0.50 and against everything written by hand; the best hand-built design came to 0.486. 144 words, 110 of them different from the ancestor's.
The first mal is a trial allocation. It fails when less than
nothing is requested — so a place is rejected arithmetically when
its quality is less than or equal to 6, and allocationally when it
is occupied. One test, two conditions, read off the return value of a memory
operation.
The underflow into an existing error path is the sign test. The organism turned a hand into an eye.
Three things make this a finding that holds rather than an anecdote.
The mechanism predicts the number
A threshold of ≥ 7 accepts qualities 7 to 15, of which 8 to 15 are good: 8/9 = 0.889 predicted, 0.876 to 0.879 measured — on five different cards.
The coupling is shown causally, not correlationally
Switch the sensor off and the quality collapses from 0.877 to 0.321, and the number of births from 4,128 to 28.
The run reproduces bit for bit
Two weeks and a dozen changes later it delivers 0.8779, to the decimal place.
And it uses none of the tools we had built for that
PC6-B noted that the instruction set was missing a sign test. In PC7 a comparison instruction was built specifically for it. In the genome of f001 it occurs zero times. The sign test had been sitting in the error path of memory allocation since day one.
The ladder, rung by rung
What the overview above compresses, here with the wrong turn that made it visible.
1 · The missing hand
The big claim of the preceding arc ran: the ecology defeats cognition, however you build the world. It was wrong — an artefact of the fact that allocation read only the size of the child, never its place. Perception had literally nothing it could have acted upon. Give it an actuator that reads an address and the finding flips: perception wins by 10.8×.
2 · Compressibility, twice
In a periodic world the whole of the information sits in one number. A genome whose length matches the stripe period places its child exactly one period further on — perfect quality, without ever seeing. We killed the period and found the lesson again one level up: spatial correlation is compressible too. At correlation length 64 the blind placer reaches 0.763. Position is inherited, and correlation turns it into quality.
3 · More than one bit
Good stripe, bad stripe — that is one bit. A graded quality from 0 to 15
demands, for the first time, a real threshold instead of a switch. The
ceiling can be computed in advance rather than only measured:
I(X;Y) ≤ H(Y).
4 · Two topologies, one intersection
In PC7-A the quality rose with every word substitution — a cleanly uphill path. And impassable all the same: the step required was three bit flips, and mutation flips one. A filler instruction that sat one flip away was what made the road walkable.
5 · The unit is the generation
For four arcs we counted in rounds and seconds. Only a generation counter showed it: four million rounds were about a hundred generations — inside which no adaptive substitution fixes, however good the gradient. And the division cost, introduced so that redundancy would persist, took a factor of 2.7 out of that. Unnoticed, because the number did not exist.
6 · Weathering — and a correction to ourselves
A precursor that does nothing yet is neutral, and what is neutral is taken apart in about three hundred generations. We wrote: we built a dial, and evolution took it off before it turned it. When we read f001, the adjustable range was still there — and load-bearing. What had vanished were two of the four anchor words our detector was searching for.
The adjustable range survived. Our search pattern did not.
7 · One verb
For seven arcs we asked why perception would not couple to action — while the entire action space consisted of exactly one action: where do I put my child. With one possible action, coupling perception to action is close to a contradiction in terms.
8 · No alphabet for messages
The runner-up exists, is one bit flip away — and nobody uses it. Not because it is unreachable, but because a signal has no addressee as long as looking is cheaper than reading. And even then the alphabet is missing: in living code all 32 instructions occur. Whoever reads a 15 cannot know whether that is a message or the decrement instruction from a neighbour's copy loop.
The gaplessness of the instruction set makes mutation harmless — and makes messages impossible.
An alphabet in which every symbol already means something as an instruction has no room for a symbol that says: this here is a message. What is missing is not a capability but a distinction. That is the wall this arc ends at.
The channels we did not build
Across the whole arc the population did better than it was allowed to — nine times. Nine times we found the reason and closed it off as a disturbance. Only while closing the arc did we recognise that they are not noise.
The working document names as one of its core questions: where exactly does the boundary between self and world lie, and through which channels does it couple? We built a sensor and assumed that was the coupling surface. Every blind route was the proof that it is wider.
| Arc | coupling exploited | through what |
|---|---|---|
| MV4 | inherited position ↔ gradient | smoothness is compressible |
| PC2-C | genome length ↔ world period | phase locking, 0.78 blind |
| PC5-B | genome length = stripe period | 1.000 blind, without a single sensor call |
| PC6-B | own position ↔ child position | block narrower than the genome |
| PC15-B | position ↔ neighbourhood quality | spatial correlation, 0.763 blind |
| PC15-C | the physics senses on their behalf | movement as “best of three samples” |
| PC4 | copy loop ↔ own output | tandem duplication for free |
None of these channels was designed as a channel. All of them were merely
not forbidden. And the sharpest is the winner's: mal is an
actuator, and f001 uses its failure mode as a sense organ.
A population under selection is a detector for coupling surfaces the designer does not know about. It has an uncomfortable property: it reports only when you have been wrong.
Why this is still not cognition
The intentional reading is seductive, so explicitly: against us assumes an intention where there is none — there was no opponent, we were wrong about the gradient and selection followed the real one.
None of these routes is an adaptation within a lifetime. All of them arose over generations. Cognition means that an organism changes its behaviour according to what it perceives now — f001 does that, which is why it is the positive finding. That is the dividing line the whole project is about.
And the touchstone: if finds structure we did not intend counted as cognition, every overfitted model would be cognitive. What remains is more precise: it is not that the organism works against us — the coupling surface is larger than our idea of it.
How the work was done
The methodological yield is at least as solid as the substantive one, and it is the part I set most store by.
Coupling is not an observation
Three metrics, three failures, one mistake in three costumes. All three are observational and try to infer perceptual use from correlation — but the question is causal.
| Metric | asks | fails on |
|---|---|---|
| Signal use | How often was something good sensed before placing? | claims coupling where there is none — geometry |
| Coupling contrast | Does it place more often after good than after bad? | denies coupling where there is some — steering is not gating |
| Control information | I(sense ; target) | tautology and perfection at once |
Two independent side findings came out of this. An organism that always places but always aims well has a contrast of exactly zero — the metric was measuring gating, while the winner was steering. And a perfect controller has zero mutual information with its own outcome, because the outcome no longer varies.
Coupling is a lesion finding, not an observation. Switch a word off, measure again.
Four habits, dearly paid for
What stays honest
So that the claim does not grow larger than the measurement.
Where it can go
All three levers that leave the substrate alone — world shape, costs, action space — have been surveyed. What is left reaches into the physics itself. That is why the arc is closed and not merely interrupted.
What remains
We have not shown under which world conditions cognition arises. We have shown what a substrate has to bring with it before the question becomes measurable — eight positions, each met as a wall and demonstrated one at a time — and we have fully explained one case in which control genuinely arose under pure replication selection.
We spent seven arcs building tools, and the only success came from something we had not recognised as a tool.
Beside that stands a second result, recognised as a result only while closing the arc. Nine times the population did better than it was allowed to, and nine times the reason was found and closed off. Those nine reasons together are the map of the coupling surface — the answer to a question we had been looking for elsewhere, collected in the folder for disturbances.