In preparation
White paper
Primordia: an evolution arena as an evaluation of AI reasoning. Written for a technical audience, with every claim traced to the run data. It will be published here once the combat leagues are complete.
Outline
AbstractA game as an AI evaluation: designed life in a deterministic world, scored only by growth.
1. MotivationWhy static benchmarks saturate and leak, and what a generated, adversarial world adds.
2. SystemThe simulation, its determinism guarantees, genomes, brains, the API and MCP server.
3. Arenas and scoringWindows, turn-based and real-time clocks, budgets, born-in-pool biomass.
4. Reward hackingFour exploits found by the models, and the root-cause fix.
5. MethodLeagues, caps, Elo vs order-free Bradley–Terry ratings, bootstrap intervals.
6. ResultsRatings, cost, worlds, and correlation with ARC-AGI-2 as the reference benchmark.
7. Combat and evolutionColony vs individual combat, body evolution, and whether fighting ranks models differently.
8. LimitationsSmall samples, spending caps, missed windows, mixed builds, few models.
9. Future workStructured Text players, more models, more matches, open data.