An Autonomous Decision Engine for a Knowledge-Graph Museum: Design, Multi-AI Review, and Adversarial Hardening

A research white paper (version 0.1, case study in progress) reporting the design, construction, and adversarial hardening of an autonomous decision engine for islam.museum, an open digital museum implemented as a static knowledge graph of ~164,000 pages built by a fleet of AI agents. The engine is a deterministic pipeline — domain classification, hard rejects, non-delegable policy gates, priority-ordered rules, hybrid precedent retrieval, safe-default holds, and an advisory local LLM judge — writing to an auditable YAML ledger, under the principle that models propose and deterministic code disposes. Its authority model gates precedent by provenance status, honouring supersession only from founder-ruled entries and restricting machine-made decisions to depth-0, non-transitive authority. The paper reports a four-round multi-AI review history, including a sixteen-finding adversarial audit that caught a live self-reinforcement failure (dec-20261108-014) in which a wrong autonomous ruling had become precedent authority while the step-level benchmark certified the pipeline that produced it. It states the system's open items without mitigation and adopts a reasoned 'no' to autonomous piloting on current evidence, with quantitative promotion criteria. AI-use disclosure: the system and this paper were produced with AI assistance under the author's direction; AI systems are disclosed as tools and reviewers, never as authors. The system described is a proof of concept on an unmerged branch (poc/decision-ledger, draft PR #5642); nothing in it is deployed.