BU74 AAI-06 Engineering Settlement Surfaces in Agentic AI Codebase Settlement Surface, Persistent Knowledge Ledger, and Small-Ontology Closure Samples

This paper develops BU74 AAI-06|Engineering Settlement Surfaces in Agentic AI as the sixth file in the B_U-based Agentic AI series. Its central claim is that Agentic AI reliability does not come from faster generation, larger models, or more autonomous execution, but from engineering settlement surfaces that keep objects, interfaces, states, verification, feedback, cost, and responsibility within the same operational world. The paper examines three engineering samples. First, the codebase settlement surface shows why bad code becomes more expensive in the AI era. Shared design concepts, ubiquitous language, stable interfaces, deep modules, static types, tests, and fast feedback are reclassified as the same-world settlement surface of a codebase. When this surface holds, AI-generated changes can be verified, absorbed, rolled back, and refactored. When it breaks, faster generation accelerates residual propagation and entropy. Second, LLM Wiki is reclassified as a persistent knowledge ledger within a specific small-ontology window. It improves upon temporary RAG by shifting from retrieval–answer–forget to ingestion–synthesis–maintenance–contradiction detection–update. Its value lies in Readout, Learning, and partial Stability closures, but it does not constitute a complete closure ecology. It maintains knowledge state; it does not automatically deliver real-world action, cost bearing, external feedback, responsibility, or full production standing. Third, vertical-domain small-ontology closure samples show that rules, tools, confidence gates, cross-validation, and narrative readouts can reduce general-model residuals inside specific task worlds. The value of such cases lies in closure engineering, not in the truth of any particular domain claim. The final judgment is that engineering standing requires settlement surfaces. A codebase needs a code settlement surface; knowledge systems need persistent ledgers; vertical domains need small-ontology tool closures. Agentic AI becomes reliable only when generation is routed through boundary, verification, cost, feedback, and settlement.