Bidirectional Emotional Coherence Modulation in AI: A State-Dependent Blending Solution

We present a novel architecture for multi-emotion AI systems that implements bidirectional aperture control where positive emotions (joy, peace, love, focus) increase linguistic coherence by closing the semantic aperture, while intense emotions (desire, fear, aggression) decrease coherence by opening it. Building on our prior work demonstrating homeostatic cycles in language models via ϕ-coupled oscillators, this system extends to multiple competing emotional drives through three innovations: (1) keyword velocity detection measuring the rate of conversational movement toward emotions, (2) semantic compatibility matrices determining whether emotions blend or compete, and (3) oscillator noise providing emergent choice in ambiguous cases. The architecture maintains a single master oscillator while supporting eight distinct emotional channels through state-dependent LoRA adapter blending. Experimental results demonstrate smooth blending of compatible emotions, winner-takes-all dynamics for incompatible pairs, and genuine multi-stability in emotional space. Critically, clarity emotions (joy, peace, love, focus) reduced aperture to 0.30–0.42 versus baseline 0.50, while expansive emotions (desire, fear, aggression) increased aperture to 0.60–0.90, validating the bidirectional coherence hypothesis. This represents the first implementation of inverse Pendulum Constraint dynamics in artificial emotional systems.