Governed Research State for AI-Assisted Analytical Work: An Implemented Reference Architecture for Evidence-Bearing State Around Frontier Models

AI-assisted research is often described as a sequence of search, reading, synthesis, and writing.Evidence-sensitive analytical work is less linear. Information arrives unevenly, questionsnarrow, and later evidence can change the relevance or standing of earlier work. Model capacityis growing, but human attention and judgement remain scarce. These conditions recur inconsulting, advisory, policy, regulatory, and diligence work, where an analysis may need tosurvive reopening, handover, or challenge after much of its rationale has been compressed into afinal artefact.This paper describes the Epistamate Research Workspace, a local-first prototype that keepsevidence-bearing analytical state outside individual frontier-model operations. The architectureseparates source identity from representation, preserved evidence from model-facing attention,model proposals from researcher-authorised state, and analytical state from reader-facingoutputs. It also records required consideration, review priority, admission, synthesis accounting,unresolved work, and researcher judgements of sufficiency for a stated use. None of thesemechanisms is claimed as individually novel.Two author-run cases exercise the current build. In a regulatory case, a contractual 48-hourincident-notification term missed by an earlier selective analysis was surfaced in a fresh run thatused a later build, a different structural representation, and full represented-unit processing, andit survived into synthesis. This is one known-positive recovery, not a recall measure. In anautomotive case, researcher-owned evidence was required for consideration without gainingelevated authority, and later context led the researcher to reject an earlier model proposal thathad become stale.The claim is bounded. The paper describes one implementation of persistent, inspectableanalytical state around replaceable model operations. It does not establish improved researchquality, retrieval quality, productivity, or human judgement, and handover between researcherswas not tested. Even with effectively unlimited model context, a system can preserve evidence,rationale, and unresolved state and order items for review; it cannot make expertise andjudgement uniform across people.