operator-control

This paper investigates semantic context drift in reasoning-class LLMs within hybrid decision support systems, proposing a mathematical model and a stability metric. A two-month experiment reveals latent goal-targeting drift and formulates engineering recommendations for control stability.