Volume or Coupling? A Scale-Dependent Dissociation in Constraint Recovery of Language-Model Loops

A language model paired with a partner is often credited with recovering after it breaks a rule, but a partner also adds generations, and a model that receives compliant text can comply by repeating it. The submitted version of this paper compared a coupled pair of Qwen2.5-Instruct instances with single instances given one, two or four generations per exchange at 1.5B, 3B and 7B, after a polite request had broken an all-capitals rule. A call-matched single model came close to the pair at 1.5B (37% against 43%), only the pair recovered at 7B (8/30 against 0/30, exact McNemar p = 0.0078), and we described a scale-dependent dissociation. Here we test that description. The difference between the two sizes' coupled advantages is 20.0 points (95% interval −13.3 to 53.3; Fisher test on discordant directions p = 0.055), the 1.5B interval reaches +28.4 points, and the 7B contrast falls to 5/30 against 0/30 (p = 0.0625) under the pilot's five-exchange window or a stable-recovery rule. The volume control and the 7B gap hold up, while the dissociation remains a hypothesis. A controlled 7B run that saves every output then asks where the gap comes from. From a common failed state, the perturbed agent keeps the rule through rounds 10–12 on 9/30 prompts with a partner, on 0/30 when the partner receives the same conflicting request (p = 0.002), and on 30/30 with a fixed reminder and no partner (p = 1.9 × 10⁻⁶, Holm-adjusted). In 20 of 21 returns to the rule, the agent copies the partner's capitalized message character for character.