Structure-First Interpretability: Anchoring Sparse Autoencoders to the Daopology Knowledge Base

Sparse autoencoders (SAEs) have become the dominant tool for decomposing large language model (LLM) activations into interpretable features. Yet the field faces a fundamental crisis: there is no ground truth for what concepts a model actually uses, and a growing body of evidence shows that SAE-learned dictionaries point in substantially wrong directions, that SAE explanations suffer from descriptive collision at scale, and that SAE features fail to improve downstream tasks beyond simple baselines. This paper argues that the root cause is the absence of an external semantic reference frame—SAEs operate without an anchor. We propose a methodology that anchors SAE research to Daopology, an axiomatic system that provides 79 minimal expressive parameters for structuring knowledge across seven categories: 9 primitives, 6 Heng families, 6 Jiao families, 8 Cha families, 7 Que families, 10 Suan families, and 21 He families. We further introduce GlyphCode, a universal symbolic encoding system that renders structural entities as compositional symbol trees, and the GlyphCode Reasoner, a neuro-symbolic inference engine that operates on these trees. The proposed “structure-first” paradigm reverses the current practice: instead of asking “what does this SAE feature mean?”, we ask “which Daopology structure does this SAE feature align with?” We outline four concrete interventions—structural initialization, structural regularization, causal verification via stability criteria, and cross-model alignment via kernel isomorphism—and provide a set of falsifiable predictions. The paper positions itself as a methodology contribution, arguing that the path forward for mechanistic interpretability lies not in scaling SAEs further, but in providing them with the semantic scaffolding they currently lack.