Toward Path-Constrained Routing in Mixture-of-Experts Models: Extending OPT as an Operational Routing Framework

Recent work has suggested that routing mechanisms in modular large language models may be interpreted as approximations of cognitive pathway selection processes. Within this perspective, the Ontology–Process–Trajectory (OPT) framework provides a structural account in which reasoning emerges from the propagation of signals between generative sources and stabilization sinks. Under such a formulation, expert routing in Mixture-of-Experts (MoE) architectures can be understood as an instance of pathway selection within a distributed computational system. However, existing interpretations remain descriptive rather than operational. In current architectures, routing decisions are implemented through statistical gating functions learned via gradient-based optimization, without explicit structural constraints corresponding to pathway dynamics. As a result, while pathway-based descriptions offer a useful interpretive lens, they do not directly influence how routing is computed within the system. This paper extends the OPT framework by introducing the concept of pathway-constrained routing. Instead of treating pathways as post-hoc descriptions of routing behaviour, pathway structures are incorporated directly into the routing process itself. In this formulation, routing decisions are defined as probability distributions over source–sink pathway units, and expert modules are aligned with stabilization mechanisms within this pathway space. Routing is therefore reinterpreted as the selection of pathways rather than the selection of expert indices. The present work does not introduce a concrete training algorithm and does not provide empirical validation. Rather, it focuses on defining the structural components required to transform pathway-based interpretations into routing mechanisms, including the construction of pathway spaces, the definition of pathway generation functions, and the formulation of pathway-constrained routing distributions. A minimal implementation framework is outlined to demonstrate how such a formulation could be integrated into existing architectures. Taken together, this work suggests a possible transition from unconstrained statistical gating toward structurally interpretable routing, and outlines a direction for developing modular systems in which signal propagation pathways are explicitly represented as part of the model’s internal computation.