Veralume : A Meta-Control Framework for Regulating LLM Responses under Semantic Tension

Large Language Models (LLMs) often exhibit unstable behavior when confronted with prompts involving paradoxes, self-reference, or high semantic tension, leading to forced resolutions, hallucinations, or incoherent explanations. We introduce VERALUME, a lightweight meta-control framework designed to regulate LLM response behavior without modifying model parameters. VERALUME operates as an external control system that estimates semantic tension in user inputs, dynamically selects a response regime, and adapts control parameters through feedback based on measured response characteristics. We formalize VERALUME as a discrete-time dynamical system with continuous state variables representing mobility, elasticity, semantic tension, and response chirality. This work makes no claims about internal cognition or consciousness; instead, it provides an analyzable and reproducible method for steering LLM behavior under epistemically unstable conditions.