Deterministic Ontological Routing Framework for Domain-Restricted LLM Systems: Design and Implementation of DROS v7.3 (確定性本體路由框架用於領域受限的 LLM 系統:DROS v7.3 的設計與實現)

We present DROS (Dharma Reasoning Operating System) v7.3, a serverless flat-file AI governance framework centered on Deterministic Ontological Routing (DOR), designedto address cross-domain semantic drift and hallucinated citation in high-precision knowledge systems. Conventional RAG systems rely on vector similarity retrieval whose inherent context blindness leads to severe knowledge contamination in domains with strict ontological boundaries (e.g., classical religious philosophy, legal texts). Drawing design inspiration from the Tang dynasty Panjiao (Doctrinal Classification) knowledge architecture — a systematic solution to apparent textual contradictions developed by Zhiyi of the Tiantai school in the 6th century — DROS shifts domain boundary enforcement from soft prompt engineering to OS-level physical path Access Control Lists (ACL), achieving zero-overhead semantic isolation. In a pilot study using 7 high-confusion cross-sectarian benchmark questions, 7/7 questions exhibited no observable cross-sectarian contamination under the current evaluation protocol; token consumption was reduced by 72%, and retrieval latency dropped from 420ms to 12ms. Given the limited sample size, this evaluation does not allow claims of universal correctness, but demonstrates a systematic improvement trend with architectural reproducibility.
-----------------------------------------------------------------------------------------------------------------------------------------

本文提出DROS(Dharma Reasoning Operating System )v7.3,一個以確定性本體論路由(Deterministic Ontological Routing )為核心的無伺服器純文字領域AI 治理框架,旨在解決高精度領域知識系統中的跨領域語義漂移(Semantic Drift )與幻覺引用(Hallucinated Citation )問題。傳統RAG 系統的「語境盲目性(Context Blindness)」在對概念邊界有嚴格要求的領域中(如古典宗教哲學、法律文本)會導致嚴重知識污染。DROS 的設計靈感來源於中國天台宗的判教(Panjiao)知識分類體系,將領域邊界從軟性提示詞工程轉移至作業系統物理目錄的訪問控制(ACL)。在以7 道高混淆度跨宗派問題為基準的先導性評估(Pilot Study )中,7/7 道題目未觀察到跨宗派語義污染;T oken 消耗降低72%,檢索延遲從420ms 降至12ms。受限於先導研究樣本規模,本評估不主張對普遍正確性作出結論,但呈現出具有架構可重複性的系統性改善趨勢。