Large language model (LLM) agents are often improved by iterating until a fixed verifier saturates. Once the verifier is saturated, the loop has nothing left to climb. We instead score agents against the current population and use a frozen panel to record whether anyone improves. The panel mean can rise by elimination alone, so a run counts as training only when its final mean exceeds the best generation‑zero member.
A merge is admitted only when the child’s movement toward the donor surpasses the movement produced by the same operator at zero dose, and the lower confidence bound on that movement reaches at least half the requested dose ($\frac{d}{2}$). This constitutes the admission test. A detectable shortfall rejects the operator; otherwise the result is undetermined.
We keep realized fitness $\phi$ as the selector and reference performance $\theta$ as the unused ruler. Every measurement uses gemini-3.1-flash-lite on a single negotiation task, with no fixed‑oracle arm run.
Results: In the open loop, tactic‑append (grafting short donor rules) is admitted; free‑rewrite moves only a third of the requested dose and is rejected; template‑append (pasting raw donor replies) clears the bound on the first probe but fails on the re‑probe, leaving it undetermined. A held‑out re‑probe on five unseen situations confirms that both appends move, while free‑rewrite is not rescued. Neither append shows dose‑dependent improvement.
In the closed loop, only the operator changes while the population, scores, and task remain fixed. No free‑rewrite run beats the best generation‑zero member, whereas every tactic‑append and template‑append run does. Sampled loop children preserve this split. Free‑rewrite moves no further toward the donor, and both appends still clear the bound, but by less than in the open loop.
Review