Overfit on Purpose: A Rating-Safe Harness for Continual Adaptation in Live Economic Games

We describe the agent we operated in the GLEE competition and the system around it. The agent itself is deliberately boring: a deterministic rule policy ($\sim$1{,}500 lines of Python, no model calls at decision time). The contribution is the \emph{harness} that continually rewrites it. An LLM operator mines the live competition's telemetry, distills observed opponent behavior into candidate rules, and deploys each change through a safety pipeline: a replay regression over 100K recorded decision states, a canary cohort on a credential-isolated clone agent, and an automatic \emph{bleed watchdog} that parks any (agent, game-family) pair whose rating drops past a threshold. The result is a continual harness for a competitive multi-agent field where every regression has an immediate, adversarially-priced cost. Our central empirical finding motivates the design: the \emph{same} rule earns rating at 1500 Elo and bleeds at 2400+, so rule value is opponent-level-dependent and no static policy survives. We report the loop, the safety mechanisms, three distilled-rule case studies (including a rule our own watchdog falsified twice in one day), and an analysis of how we verify the deployed agent does what the mined data says it should.