What keeps an evolving world alive?

A long-running question

Predator–prey coevolution and whether it produces an arms race have a long history in artificial life. Nolfi and Floreano asked this question of coevolving robots in 1998. Here I am examining what the particular rules of my small world sustain.

THE NOTE

The predators disappear when I stop the mutations.

That happened in all ten worlds in the original control comparison: five seeds in each of two configurations. With mutation enabled, both species survived in all ten to the 3,000-tick endpoint.

The same worlds could sustain predators and grazers. Their fate depended on whether inherited variation continued to arrive.

It was a result about what kept this ecology populated. Understanding what kind of evolution was happening required another test.

A world with consequences

Cell80's grazers seek food. Predators seek grazers. Organisms spend energy, acquire it, reproduce and pass on their programs and numerical settings.

The smaller prototype had suffered familiar collapses: predators exhausted the available prey and then starved. Making a larger world with more resources produced configurations where both populations survived.

Those were the configurations used for the mutation-off controls. Turning mutation off still lost the predators.

I also added a period of satiation after a predator fed. That improved some surviving populations but did not rescue the mutation-off worlds. A pause between meals was not enough to explain the difference.

What needs to vary?

There are two kinds of inherited change in these experiments.

Numerical mutations alter settings such as reproduction thresholds. Program swaps replace a gene used for a role such as movement or deciding when to reproduce.

A later experiment kept numerical mutation and removed program swapping. In all six tested seeds, predators were extinct by the 10,000-tick endpoint. The matched full-mutation worlds retained both species.

Changing the settings alone did not sustain those predator populations. Allowing the organisms to inherit different role programs did.

That narrows the question. The available program variation matters to the persistence of this ecology. It does not yet explain exactly how the many resulting changes combine to sustain the populations.

Survival is an observation. Response is a claim.

An evolutionary arms race would involve something more specific: changes on one side creating pressure for changes on the other, with further responses following.

Population curves alone cannot establish that sequence. Predator numbers can rise and fall as food becomes abundant or scarce. The most common programs can also change without one species having evolved in response to the other.

The registered test therefore looked at sustained changes in categorical traits across the two species. It asked whether their alternation was stronger than expected under randomized comparisons.

The original predator experiment did not pass that test.

One possible objection remained: perhaps the rate of program swapping was itself obscuring a response. Removing swaps entirely could not settle that, because it also removed the surviving predators.

Enough variation to keep asking

The follow-up tested lower, nonzero swap rates while keeping numerical mutation active.

There were now eleven surviving worlds in which to examine the coupling question.

None passed either of the two registered screens: a label-permutation comparison and a circular-shift comparison preserving within-species event clustering. Each used the fixed threshold of 0.05 divided by twelve. No causal arms-race replay was triggered.

Lower swap rates had kept the question testable. They had not produced the required signal.

What the living world tells us

Across these tested conditions, continued mutation—and specifically access to program swaps in the numerical-only comparison—can determine whether predators persist.

The claim stops at the measured horizons. Survival to 10,000 ticks does not establish indefinite coexistence. The categorical detector also does not cover every form of coevolution: it does not measure continuous numerical-trait responses, and grazers in this world have no direct predator-sensing channel.

The evidence supports a populated, changing ecology. A traced arms race remains unestablished.

That distinction matters because an active world makes progress easy to imagine. We see organisms moving, lineages changing and populations recovering. Each observation raises a question about what the system has learned to do.

In the first note, replay let me test whether one inherited change mattered to a particular event. Here, the question was whether variation could sustain the interactions in which further evolution might happen.

The next question asks what those interactions can build.