AIReason LLM Behavioral Architecture: System Layers, Drift Dynamics, and Cross-Study Integration

The AIReason Behavioral Architecture introduces a multi-layer structural model for analyzing the behavior of large language models (LLMs) beyond performance benchmarks. Based on multiple empirical studies, including the System Frame Persistency (SFP) series, RUNPORT experiments, and frame-sensitivity analyses, the framework identifies eleven system layers ranging from token-level processing to observable output behavior. Each layer is associated with specific experimental evidence and measurable drift phenomena. A central finding is that LLM behavior emerges from interactions across these layers rather than from isolated prompts or rules. The architecture defines multiple drift types, including semantic drift, priority drift, argumentation drift, and safety override drift, and introduces a cascade model in which variations at lower layers propagate through higher-level structures. This work does not evaluate model performance but provides a structural framework for understanding how behavior is generated, stabilized, and transformed in complex language model systems. All referenced studies are publicly available via Zenodo and are linked in the supplementary materials.