LLMs as Constrained Translators: Ontology-grounded Multi-hop Reasoning from Traditional Chinese Medicine Syndromes to Traceable Molecular Evidence

Traditional Chinese Medicine (TCM) encodes combinatorial diagnostic logic in specialized syndrome language (zheng) that resists recovery by general-purpose NLP. We propose that LLMs are most effective as constrained translators of expert language---mapping syndrome text to ontology vectors---while deterministic pipelines handle evidence retrieval. We formalize this as a role-division principle and present TCM-HOP, a five-hop weakly supervised inference chain that exploits implicit signals already present in clinical case records, requiring zero direct syndrome-to-compound annotations. The framework achieves R@3=0.617 with 0% hallucination, covers all 42 reference genes for diabetic nephropathy, and earns 93.3% expert path retention. Applying the framework, we operationalize syndrome-conditioned treatment differentiation as a measurable NLP task (permutation p=0.001) and uncover systematic diachronic language drift across centuries of TCM practice (P@5: 0.588 vs. 0.137, p<0.001). More broadly, TCM-HOP maps the boundary between where LLMs serve TCM and where they fail, and shows how its syndrome-conditioned logic can be made measurable.