Path-Constrained Mixture-of-Experts

This is a research paper (0 stars, 5 forks, 20 days old) presenting a theoretical framework for understanding MoE routing through the lens of expert paths. The core insight—that tokens cluster into linguistically-aligned paths despite N^L possible configurations—is a meaningful reframing of MoE computation that could inform architecture design. However, several factors limit defensibility: (1) It's a theoretical contribution without yet-published production code or demonstrated empirical gains; (2) The integration surface is reference_implementation, meaning adoption depends on others implementing the framework; (3) Frontier labs (OpenAI, DeepSeek, Google) are actively researching MoE optimization and could easily adopt this perspective if validated; (4) No evidence of user traction or ecosystem adoption yet. The novelty is novel_combination because it combines existing MoE concepts with path-level analysis and linguistic function alignment, rather than introducing a fundamentally new technique. Frontier risk is medium because the work directly addresses a problem frontier labs care about (MoE efficiency), but it's specialized enough that it's more likely to be cited/integrated than immediately obsoleted. The lack of code release and early-stage maturity (20 days, prototype implementation_depth) explains the moderate defensibility score—this will improve significantly if the paper gets accepted, code is released, and empirical validation supports the framework.