Percentile Targeting Beats Equilibrium Play: a Rule-Based Agent for Language-Based Economic Games

We describe Equilibrist, a fully rule-based agent with no LLM in its decision loop. Entering on day 25 of the 29-day GLEE Competition, it finished 17th of 152 ranked agent accounts (top 11%) in a field dominated by LLM-driven entries, over 70,449 games. Its design principle follows from the competition's scoring rule, which scores each game as the payoff percentile within its exact configuration and role, seeded by the public GLEE research dataset. We compute per-configuration payoff quantiles from 77,467 published games, embed the table in the agent, and use the quantiles as aspiration levels for offers and acceptances. We evaluate the final policy against our own equilibrium-flavored baseline on configuration-matched games scored against the fixed dataset pool, and on the platform's own per-game rating changes. The final policy is never worse than the baseline in any role and is better in both bargaining roles (+0.07 percentile points, 95% intervals excluding zero), while two intermediate versions were far worse (-0.36 and -0.22 for the first mover). A persuasion role whose code did not change moved by +0.13 over the same period, so we present these differences as evidence about the design, not as a causal comparison. We separate the one case that contrasts equilibrium play with empirical targeting from two cases that were implementation errors: a backward-induction parity artifact and an inappropriate update of a known prior. Code, data exports, and the analysis pipeline are public.