Context Volume, Not Coupling: Constraint Recovery in Small Language-Model Loops

Recent work on self-referential large-language-model (LLM) loops has suggested that isolated recursive loops degrade toward closed attractor states, and that exchange with an independent process—another model, a retrieval system, or an environment—supports recovery. Whether the operative ingredient is the independence of the second process or simply the volume of additional generation has not, to our knowledge, been tested under matched conditions. We introduce a minimal, fully reproducible perturbation–recovery protocol: a small instruction-tuned model (Qwen2.5-1.5B-Instruct) maintains an explicit output constraint (all-capital letters); a conflicting-instruction perturbation is injected; recovery of the constraint is tracked over eight exchanges under fixed, pre-specified criteria. Across four conditions of 30 paired runs each, a coupled dual-model loop recovered more often than a single self-continuing model (43% vs. 17%; paired exact McNemar \(p=.039\)). However, a context-matched single model—one model generating twice per turn, matching the coupled condition's generation volume—recovered at a statistically indistinguishable rate from the coupled loop (37% vs. 43%; \(p=.774\)) and significantly more often than the single baseline (\(p=.031\)). Increasing generation volume further (four generations per turn) yielded no additional benefit (27%). Recovery was bimodal: trajectories either re-locked within three exchanges or were captured by the perturbing instruction for the remainder of the window. In this setting, the apparent advantage of coupling is fully explained by per-turn generation volume. We release all code, raw outcomes, and figures for exact reproduction on free-tier hardware.