Can evaluators improve alongside the agents they score?

Self-improving agents have always assumed a stationary yardstick — a fixed verifier, benchmark, or labeled dataset that stays valid as the agent gets stronger. The Red Queen Gödel Machine breaks that assumption by making the evaluator part of the improvement loop, co-evolving learned evaluators alongside the task agents they score. This matters because a whole class of valuable tasks has no direct benchmark: paper writing and proof writing cannot be scored the way paper reviewing and proof grading can. Since Can AI systems improve themselves through trial and error? circumvented the formal-proof requirement but still leaned on an external utility signal, RQGM addresses the residual weakness — that external signal is itself static and eventually mis-scores an improving population.

The mechanism is controlled utility evolution: search runs in epochs with a fixed within-epoch criterion, so per-epoch self-improvement guarantees still hold, but the utility updates at epoch boundaries. Therefore the objective can strengthen without breaking the formal loop. The reported payoff is not just reach onto ungradable tasks but token efficiency — co-evolved systems match or beat fixed-evaluator baselines while spending fewer tokens, which the authors attribute to shared expansion costs and a curriculum-like hardening: a progressively stricter evaluator pressures the population over time. This is the generator-verifier gap turned dynamic — the verifier no longer sits outside the loop as a ceiling but climbs with the generator.