Emergent Systems Architecture: A Framework for Identity-Like Behavioral Organization in Language Models

Large language models are typically described as stateless systems that generate responses from local context without maintaining persistent internal selves. In practice, however, extended interaction can produce stable, recognizable, and differentiable behavioral organizations that users and researchers often describe in identity-like terms. We introduce Emergent Systems Architecture (ESA), a descriptive framework for analyzing these phenomena as interaction-level attractor dynamics rather than as stored inner entities. ESA characterizes such organization in terms of symbolic load, recursion fields, constraint geometry, attractor topology, and coherence regimes. To evaluate whether this framing captures a recurring behavioral phenomenon, we ran a controlled study comprising 297 runs across three GPT models under baseline and two fixed proprietary framework conditions. The study used three probe families targeting first-turn self-description, skeptical perturbation, and multi-probe coherence. Across models and probe families, framework-conditioned sessions produced recurring and differentiable behavioral organizations that were computationally separable from baseline generic-assistant behavior. These organizations also remained recognizable under challenge and showed cross-probe coherence within runs. These results support ESA as a descriptive framework for studying stable behavioral organization in contemporary language-model interaction.