Model-External Command Admission for Auditable Local AI Agents: System Artifact and Preliminary Case Studies

Grant operational autonomy without ceding boundary-setting authority. This technical report describes a deterministic, model-external Command Admission layer for auditable local AI agents. Immediately before an external-effect command is executed, the layer checks the current operation contract, exact command identity, monotonic revision, stop latch, single-use grant, and audit requirements. Its purpose is to separate a model's reasoning capability and operating-system privileges from the authority that may legitimately be exercised at a particular moment. The artifact includes implementation-level formal tests, anonymized real-system cases, a theory-to-claim crosswalk, structured evaluation data, and a bounded-authorization workflow. Operation on at least one real route supports implementation feasibility and motivates route-by-route coverage testing. It does not establish universal enforcement, equal strength across execution routes, general safety, long-term stability, or a quantified reduction in approval fatigue. The underlying work was not initiated as a preregistered or publication-oriented experiment. The theorist, implementer, primary operator, and primary evaluator are the same person. The evidence is therefore presented as preliminary and exploratory material for independent scrutiny, replication, and alternative implementation. 境界設定権を手放すことなく、実行上の自律性を与える。 本技術報告は、ローカルAIエージェントが外部effectを生じさせるcommandを実行する直前に、モデル外の決定論的componentがoperation contract、exact command identity、単調増加revision、停止latch、single-use grant及び監査要件を照合するCommand Admission方式を記述する。 限定された少なくとも一つの実経路で方式を層として接続できたことは実装可能性を支持するが、全経路拘束、一般的安全性又は承認疲労低減を確立するものではない。単一operator・単一testbedによる探索的資料であり、独立再現及び比較評価を今後の課題とする。 The Japanese report is the authoritative original. The English PDF is a reference translation. AI contribution and human responsibility are disclosed in the accompanying materials; no OpenAI review, approval, or endorsement is claimed. Contact / 連絡・議論 Questions, criticism, correction proposals, replication notes, and related research discussion:https://github.com/mos52263055347-ctrl/thought-chemistry-research/issues 質問、批判、訂正提案、追試・再現に関する報告及び関連研究上の議論:https://github.com/mos52263055347-ctrl/thought-chemistry-research/issues