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

Foundation-model agents deployed in dynamic environments must maintain reliable behavior under changing conditions and unexpected perturbations. However, apparent recovery after failures may result from increased generation capacity rather than genuine restoration of stable behavioral constraints. We investigate this problem through a controlled perturbation–recovery framework for language-model loops under conflicting instruction constraints. Across multiple model scales, we compare coupled multi-agent interaction, single-agent recovery, context-matched baselines, and volume-controlled generation conditions. Recovery behavior is evaluated through deterministic replay of interaction trajectories. Our results reveal a scale-dependent dissociation between interaction structure and generation volume. At smaller scales, additional generation budget explains a significant portion of apparent recovery improvements. At larger scales, coupled interaction produces recovery patterns that cannot be reproduced by increased token budgets alone. These findings highlight the importance of evaluating foundation-model agents beyond static performance metrics. Our framework provides a methodology for studying behavioral persistence, robustness, and recovery mechanisms in adaptive agent systems.