Evidence Authority After Retrieval: A Control-Surface Framework for RAG and Agentic Systems

Abstract

A prior LatentAtlas methodology preprint argued that relevance is not authority: a retrieved source may be topically correct while still being insufficient for action, publication, identity resolution, or customer-facing use. Building on that result, this paper develops the post-preprint thesis that evidence authority is an operating control surface between retrieval and downstream use. The proposed layer receives candidate sources from retrieval, binds them to an explicit claim and use context, tests provenance, freshness, identity scope, contradiction state, source role, support axes, and intended use, and returns a controlled lane before generation, action, or publication proceeds. The paper introduces a minimal evidence authority packet, separates retrieval, qualification, generation, action, and audit planes, and argues that review, learning, and release-readiness must remain governed states rather than implicit permission. The contribution is a design framework for RAG and agentic systems that makes retrieved context operationally usable without allowing similarity scores, plausible grounding, or post-hoc review labels to become unexamined authority.