Two species of neural network — one hunting, one fleeing — evolving against each other in real time, under the law of the Red Queen.
In 1973 Leigh Van Valen proposed the Red Queen: in a world of coevolving rivals, you must keep running just to stay in place. Primordium puts that idea on a track you can watch.
A predator and a prey share a wrap-around arena. Neither is programmed to hunt or to flee — each is steered only by its own artificial neural network, and those networks are shaped by nothing but mutation and selection across generations. The predator is a touch faster; the prey a touch more nimble. Every improvement on one side is a new problem for the other.
The question the experiment actually tests is sharper than “who wins”. It asks whether the arms race cycles — whether strategies chase each other round in a loop the way real predator and prey do — or whether it just grinds to a stalemate. And it answers with a control, replicate runs, and a statistic, not with a vibe.
The controllers don't start clever. Every capability below has to emerge from the genome — and across ten runs, each one reliably did.
Networks update in discrete time. Any connection that points backward — or loops to itself — carries the last step's activation forward, giving the agent genuine internal state. Recurrence emerged in 10/10 runs.
Every synapse carries an evolvable learning rate eta. During an encounter, weights are nudged by reward-modulated Hebbian plasticity — the agent learns within its own lifetime. The rule is inherited; the learned weights are not.
NEAT-style mutation adds nodes and connections under a global innovation registry, with speciation and fitness sharing. Brains start minimal and complexify only when complexity pays. Hidden nodes appeared in 9/10 runs.
No training data, no gradient descent, no objective hand-written beyond “close the distance” / “open it”. Just a tournament repeated thousands of times.
Each agent reads the relative bearing, distance and motion of its rival across the toroidal world.
Networks drive turn & thrust over a timed chase. Payoff is continuous and zero-sum — no saturating ceiling.
Every individual is tested against a random sample of the opposing population, so no one overfits a single foe.
Fitness-shared selection, mutation of weights, learning rates and topology — then the next generation enters the arena.
One self-contained page, no build step, nothing sent anywhere. Each tab is a different lens on the same evolving lineage.
Pit an evolved predator from one generation against prey from another. Lock them to the same generation, or break the diagonal to expose the Red Queen directly.
A pre-registered hypothesis, coevolution vs a frozen-target control, replicate runs, and a 20 000-shuffle permutation test that prints a verdict — including “no effect”.
An actual evolved controller, drawn live: excitatory and inhibitory edges, dashed recurrent loops, ringed plastic synapses, node brightness tracking real activation.
A short field write-up: the mechanisms, the fitness, why this control, the two measurement lenses that can disagree — and an honest account of what the experiment is not.
Primordium is deliberately modest about its claims. Cliff & Miller coevolved recurrent pursuit/evasion networks and built CIAO plots three decades ago; this rebuilds that lineage in miniature, in a browser, so anyone can run it. Population sizes and mutation rates are hand-tuned for tractability, not derived — and the code does not tune itself toward significance.
On honest reporting. Coevolution frequently fails to produce clean cycling — collapsing into disengagement or “mediocre stable states”. A null result here is a real outcome, not a broken instrument. The verdict is whatever the data say.
Open the lab, hit New lineage, and let the first champions evolve.
▶ Enter the laboratory