Governed Research State for AI-Assisted Analytical Work: Architecture and Case Studies from the Epistamate Research Workspace

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Frontier models and retrieval systems now supply strong upstream intelligence for evidence-sensitive analytical work, but the end-to-end workflow around them is often left implicit or distributed across tools. The problem extends beyond academic research. Consulting, advisory, policy, regulatory, and diligence work can follow the same iterative pattern without necessarily being labelled research: evidence is abundant and heterogeneous, deadlines force decisions before completeness, new information keeps arriving, and poorly supported conclusions can carry real consequences. Once material enters such an analysis, recurring questions remain: which sources deserve model attention, what bears on the decision framing, what remains unresolved, what changed when new evidence arrived, and who authorises a model proposal into accepted state. This paper describes the Epistamate Research Workspace, a prototype that keeps this governed state outside individual model interactions as a persisted layer connecting evidence, unresolved questions, accepted analytical objects, and authority transitions. Its contribution is an operational contract with four separations: source identity from source representation, preserved evidence from model-facing attention, model proposals from researcher authority, and generated outputs from the analytical state beneath them. We exercise the design on EU-India Free Trade Agreement research for German and EU-headquartered manufacturers, a task requiring agreement texts, policy collections, corporate reporting, and sector regulatory gateways to be related to a defined decision framing. In the historical run, 2,168 passage objects were represented and three bounded passes exposed 67, 60, and 60 of them to frontier reasoning. A 424-page Bosch annual report produced 1,055 passage objects; the first pass selected 12 rather than sending the whole document to the model-facing operation. A later continuation began from explicit evidence obligations and six acquisition targets, represented 1,215 passage objects, and ended with four findings and four knowledge gaps. The researcher stopped because the remaining automotive tariff and conformity questions required different source acquisition rather than further reasoning over the same represented pages. These cases are existence demonstrations and defect-discovery records. They do not establish comparative accuracy, completeness, productivity, selector recall, or better human decisions, and applicability beyond the exercised cases remains an architectural argument rather than a validated deployment claim. Epistamate is positioned as complementary to upstream research and retrieval systems, with model and retrieval components remaining bounded and replaceable.