When does improvement become invention?

I wanted to see it for myself

This is an old question in artificial life. In 2003, Lenski, Ofria, Pennock and Adami used Avida to show complex computational functions building on simpler functions that had evolved earlier. I had a much smaller world with traceable lineages and replayable history. I wanted to see how far I could get.

THE NOTE

An artificial ecology can change. It can adapt. It can discover things that help.

But does anything new become the foundation for what comes next?

In Cell80, organisms carry small programs as genes. Those programs help determine when they eat, how they move and when they reproduce. Offspring inherit programs and numerical settings, with mutations. Some lineages leave more offspring than others.

In the first note, I traced a population shift to a birth and tested what happened when one inherited change was undone. The second asked what kept predators and grazers alive long enough for evolution to continue.

Here I use replay to ask a different question: did a mutation help its carrier reproduce, and did a later advantage depend on it?

I completed a batch of 379 primary worlds, with additional verification runs, to separate several claims that are easy to collapse into one story.

Change → benefit → novelty → succession → dependence

Each requires evidence of its own.

01 / CHANGE

A mutation happens.

An offspring receives a different movement program, or a different threshold for reproduction. Its genome is now different from its parent's.

This is the easiest step to produce. A changing genome tells us that variation exists. We still have to find out what that variation does.

8 of 108 tested composition changes helped.

I tested changes that replaced an existing movement gene with a composition of library programs. In eight cases, the focal organism produced more direct offspring than it did when that change was undone. Eighty-four cases tied; sixteen produced fewer.

These were local advantages, measured in particular birth contexts over a fixed horizon. They establish that some composed genes can help in this ecology.

The compositions themselves were generated before the worlds ran. Evolution encountered a prepared pool of possibilities; it did not write these programs during the experiment.

5 of the 8 helpful compositions repeated an existing behaviour.

A program can contain different code and still make the same decisions.

The movement input in this world is small enough to check exhaustively: three food readings, each either zero or forty. That gives eight possible input combinations.

Across all eight, five helpful compositions chose exactly the same immediate actions as a gene already in the library. Even the composition with the largest offspring gain repeated an available behaviour.

Three helpful compositions had distinct action signatures in this domain. That is a specific, bounded kind of novelty. The check concerns immediate movement actions; it does not establish equivalence of execution costs or future mutation possibilities. It was also a later diagnostic, separate from the registered acceptance tests.

2 candidates contained a later beneficial change.

I followed descendants that retained a helpful composition and tested subsequent numerical mutations. Of 72 tested changes, two increased the focal organism's direct offspring count.

One altered how much energy a parent passed to its offspring. The other lowered the reproduction threshold.

There was now a sequence worth investigating: a helpful composition, inherited along a lineage, followed by another helpful change.

Getting better twice raises a further question. Did the first improvement contribute to the value of the second?

Neither candidate passed.

For each candidate, I tested four combinations in fresh worlds: the original genes, the first change alone, the second change alone, and both changes together.

The registered test asked whether the later mutation helped more with the composed movement gene than with the original movement gene. That interaction had to be positive in at least four of five held-out worlds, and positive on average.

One candidate had positive interaction in three worlds. The other had it in none.

The assay used whole founder populations in fresh environments, so it answers a different question from replaying a single birth in the ancestral world. Under that assay, neither sequence established the registered dependence claim.

There is also a stronger question beyond this test: would the later mutation have been advantageous at all without the earlier change? A positive interaction alone would not settle that.

06 / THE RATCHET

Still open.

Getting better isn't the same thing as building on what came before.

The result I want next is an inherited capability that makes a later advance possible. A lineage crosses a barrier; something previously unavailable becomes useful; subsequent evolution builds on that access.

These experiments have not demonstrated that sequence. Even a successful bounded example would leave the larger question of open-ended evolution unresolved.

A useful next experiment would introduce a resource or environmental structure that no existing primitive can exploit, but a composition can. The primitive library would need to be checked against that barrier, and the lineage would need to acquire and retain the capability through selection.

Then comes the decisive comparison: a later mutation should help when that capability is present and fail to help when it is absent, across held-out worlds. That would test whether the first advance opened a route to the second.

The finite experiment is complete. The question it leaves is larger:

THE EARLIER RECOVERY CLAIM

The earlier numerical recovery “ratchet” is separate from cumulative capability dependence. Extending the no-swap condition from five seeds to ten lowered mean recovery from 0.505 to 0.359. No tested full-role dose reached the registered 0.5 threshold. This genome endpoint is not reproductive fitness.