Cognitive Stability as the Missing Layer in Long-Horizon LLM Collaboration

This paper reports operational findings from long-horizon theoretical collaboration with large language models (LLMs) under high-load conditions, without any modification to model architecture, training data, or inference procedures. It identifies cognitive instability—manifesting as referential drift, version confusion, state discontinuity, and compression-induced erosion of logical structure—as a dominant bottleneck in sustained intellectual work. The paper then describes systematic, human-led interaction-layer interventions that substantially mitigate these failures, including definition freezing, provenance scaffolding, explicit externalization of reasoning traces, curriculum-based introduction of novel frameworks, and controlled cross-model perturbation to escape deadlocks. These interventions enabled the completion of multiple long-form theoretical artifacts publicly archived with persistent identifiers. The paper argues that cognitive stability constitutes a missing operational layer toward reliable long-horizon research assistance and proposes minimal design principles implementable using existing technologies.