What you're looking at
Each line is a country. Its height is its fitness — how well its bundle of policies is doing right now. Every step, each country in turn flips the single policy that helps it most, if any helps at all. But the countries are coupled: one country's move re-scores its neighbours, so their best moves shift too. This is the co-evolutionary substrate from the essay — the baseline dynamics, before any language model takes the wheel.
When no country can improve by moving alone, the world has reached a weak Nash equilibrium — a truce. Hit Perturb to randomise a few policies and knock it off that truce, then watch whether it walks back.
- N
- policies per country — the length of its genome.
- K
- how many of a country's own policies each policy depends on. More tangle, more jagged the landscape.
- C
- how many policies in other countries each policy depends on. This is the coupling that makes it a game. Low C settles; high C chases its own tail forever (a "Red Queen" churn).
It defaults to numbers that settle. Push Coupling (C) up past the internal tangle and the truce stops forming — the same edge-of-chaos the landscape was built to show. The full method, and where the language models come in, is in the essay and its white paper; the code is at github.com/matbest/LLMJupyterTester.