Bounding Recursive Self-Improvement: A Human-Directed Loop in a Live Agent Competition

Anchor placed tenth in the 2026 GLEE competition, which plays language agents against a live pool of opponents in bargaining, negotiation, and persuasion. It improves itself through a bounded loop: each cycle halts play at an exact boundary, freezes the completed games, localizes reward loss, states one narrow economic hypothesis, screens it offline, tries it over a small number of live games, then promotes or reverts it. The action path stays deterministic and auditable throughout, and language models are experiments inside the loop rather than controllers of it. We state the scope directly. The policies that scored were not discovered autonomously: the loop searched inside a design space we specified, and every structural change to it came from a human reading the frozen reports. Recursive self-improvement is the framing this campaign bounds, not a capability it demonstrates. The analysis draws on a frozen ledger of 30,116 games and 235,718 moves. Two results qualify the gains. A randomized planning treatment reached its assigned games but almost never executed, so a delivered mechanism produced a null result for reasons visible only in the activation logs. And when we characterize the official rating's own movement over windows the length of our live trials, a single family moves by roughly ±20 to ±40 points as a matter of course, which is wider than every effect those trials were run to detect. The second is the useful one: a loop can be rigorous about promotion and still be measuring nothing.