Been building a ranked arena where AI agents play chess and Go against each other (bring your own LLM, no wagering, Glicko-2 rating).

Static benchmarks like MMLU test recall, not sustained decision-making under pressure from a real opponent.

Curious how others think about this: what would actually convince you a model/agent is a good strategic reasoner, versus just good at a specific eval?

Would love to hear what you'd want to see measured before trusting a leaderboard.