How this was made

This project was built by one person, in extended dialogue with several AI systems — Claude (Anthropic, Opus 4.8), Codex (OpenAI, GPT-5.5), and Fable. I want to be clear about that, and clear about what it did and did not mean, because “made with AI” now covers everything from honest collaboration to hiding the absence of work.

It was not used to write this for me.

It was used to argue against me.

The working method

The shape of the work was adversarial on purpose. I would bring a hypothesis — about what kind of environment stays stable, what a blind referee can and cannot do, why a result held or failed — and the job of the models was to try to break it. Most hypotheses did break. The ones that survived being attacked are the ones that made it into the essay and the experiments.

Concretely, the division of labour looked like this:

• Claude was the reasoning partner — used to stress-test ideas, find where an argument quietly assumed its own conclusion, surface prior work I should have known, and push on the writing until it said what I actually meant. When a claim in the essay is careful — “this looks like a limit, not a proof” — that caution is usually the residue of an argument I lost to it first.

• Codex was the engineering partner — used to build, run, and audit the simulations: the worlds, the blind referee, the validation checks designed to ensure the referee never reads hidden labels, and the thousands of seeded runs behind every reported result.

Neither model was treated as an authority. A claim survived only when it was still standing after repeated attempts to break it — sometimes by me, sometimes by the models, often by both. The two were used the way you would use sharp colleagues who do not care about your feelings: to make weak ideas harder to keep, not easier.

What stayed mine

The questions were mine, and so were the constraints. The framing — soil rather than seed, blindness as the constraint rather than a handicap — was mine. The decisions about what counted as a result, what to keep, and what to throw away were mine. And the responsibility for every claim on these pages is mine: if something here is wrong, it is wrong because I let it through, not because a model said it.

I think that is the honest way to describe this kind of work right now — not authored by AI, not done without it. Built by a person who used very capable tools to be wrong less often, and who remains accountable for whatever wrongness survives.

— Kirill Kruglov