Clause AI-2: The Gradient Starvation Envelope for Sparse MoE Training Dynamics

Executive Summary: This technical specification defines the Gradient Starvation Envelope (Eq. 114.1), a formal compliance primitive designed to stabilize the training of Sparse Mixture-of-Experts (MoE) architectures.Technical Scope: The framework utilizes a Lyapunov-style differential inequality on per-expert gradient variance to eliminate "routing collapse" and "dead experts" while ensuring training health. It provides the mathematical bridge between high-performance MoE dynamics and enterprise AI governance requirements, such as EU AI Act Article 11/53 and NIST AI 600-1.