Frontier models are getting better at finding information and producing answers. The harder problem is governing what deserves attention, what becomes accepted, what remains unresolved, what changed, and who owns the decision.
Epistamate keeps three things deliberately separate across the workflow: model proposals, governed evidence state and human authority. The same governance kernel can be mapped to research, advisory, policy and other evidence-sensitive methods without forcing them into one generic workflow.
Professional analysis can begin with client documents, internal research, enterprise RAG, licensed material, scholarly or official sources, targeted external search, or research produced by another specialist system. Epistamate is designed to build on that information landscape, preserve more than any one model packet can inspect, and make the next use of attention explicit without inheriting authority merely because an upstream tool produced a polished answer.
Research is one concrete implementation of this discipline. Advisory and other professional workflows can reuse the governance kernel while keeping their own stages, evidence types, judgement vocabulary, feedback loops and purpose-specific outputs.
These questions recur in research, consulting, policy, diligence and other evidence-sensitive work.
A supplied industry fraud/cyber paper looked well-supported on the surface. Epistamate treated each step separately instead of assuming that a working link meant a supported claim.
The run parsed 27 citations. Ten routes were reachable; 17 failed or were blocked. After preserving and de-duplicating what could actually be obtained, 11 unique source contents remained for analysis.
Route reachability, citation identity, claim support and substantive validity remained separate questions. Duplicate source content was not counted as independent corroboration, and unresolved support questions survived into the research record rather than disappearing into fluent prose.
Controlled product-validation example, not a client outcome claim or comparative accuracy benchmark.
In a separate question-led run, acquisition stopped with material still unresolved. The workspace carried knowledge gaps and validation leads forward instead of allowing a first-pass synthesis to imply completeness.
In a synthetic advisory-method experiment, accepted evidence could make linked propositions, findings, recommendations and deliverables worth reopening while the consultant retained authority over each consequential transition.
These are three current applications of one underlying governance architecture, not three unrelated products. Epistamate does not assume research, advisory and assurance should use the same nouns or stage sequence. The shared kernel governs evidence identity, bounded model work, authority, dependencies, history and change; each professional workflow keeps its own method semantics.
The most complete current expression: question-led, document-led and hybrid research with bounded model passes, claim/support review, inspectable attention, persistent gaps, iteration lineage and outputs shaped to the researcher's intended use.
Map the same governance discipline onto an existing advisory method. New evidence can redirect attention and mark downstream recommendations affected without replacing the consultant’s stages, terminology or judgement.
A bounded evidence review of one consequential GenAI workflow: what claims management relies on, what controls and operational evidence actually exist, and what the available research establishes about those assumptions.
Local-first in the current Research Workspace: canonical state, review decisions and locally supplied evidence are stored on the user’s machine; explicit web acquisition or external-model operations may send the material required for that operation to the configured external service.
Recommendations built from research, client documents, interviews and specialist input where dependencies and later changes matter.
Work where source authority, jurisdiction, applicability, disagreement and known unknowns must survive the brief.
AI-assisted work product where a reviewer, partner or client needs to inspect the basis — not only receive a confident answer.
Epistamate is not trying to replace the systems that find information or provide domain intelligence. Those can remain upstream. Its role is the governed workflow that follows: what entered the analysis, why something received attention, what became accepted, what stayed unresolved, what changed later and how that state should shape the professional output.
Use the strongest scholarly, enterprise, domain or frontier-model capability available. Epistamate does not need to own the largest corpus to govern what the workflow does with the resulting information.
Separate source identity from representation, preserved evidence from model attention, model proposals from human authority, and generated outputs from the state beneath them.
The shared state can project differently by profession: a paper, policy position, recommendation, clause/remediation priority, diligence issue list or other deliverable — without changing the evidence and judgement beneath it.
If you mainly need ideas, creative drafting or a quick answer, the extra evidence discipline may be unnecessary.
Both, but they are distinct. The Governed Research Workspace is the current software product in early access. GenAI Workflow Assurance and Consulting & Advisory are bounded service applications that use the same evidence-control architecture.
No. Epistamate is designed to use strong upstream intelligence for search, retrieval, reading, reasoning and challenge. Its role is to preserve the evidence, unresolved state, dependencies, changes and human decisions around that work.
No. The shared layer is the governance kernel: evidence identity, bounded model work, candidate-versus-accepted state, authority, dependencies, history and change. Each domain can keep its own stages, terminology, evidence rules, judgement vocabulary and output formats.
No. Epistamate does not calculate a universal confidence score or source-independence score. Directness, source role, exact support, independence, applicability, contradiction, recency and sufficiency remain separate, inspectable dimensions rather than being collapsed into one number.
It should not rely on a single opaque score. Attention can be driven by inspectable state such as contradiction, weak or sparse support, applicability uncertainty, missing documents or facts, stale downstream work, researcher priority or newly arrived evidence. The researcher should be able to see why an item surfaced, what evidence/state caused it, what it affects and what action could resolve it.
No. The durable research state is the substrate; the output is a projection for the intended use. Academic work may end in a paper or literature synthesis, policy work in options or recommendations, advisory work in a decision or recommendation set, and compliance work in an analytical brief, clause priorities or remediation actions.
In the current Research Workspace, canonical research state, review decisions and locally supplied evidence are stored on the user’s machine. Explicit web acquisition or external-model operations may send the material required for that operation to the configured external service. Local-first does not mean zero-network.
The architecture is designed around local research state and bounded external assistance, but you should not send confidential material to an external model unless your own policy and provider terms allow it. Closed-source and more tightly bounded workflows remain an important product direction.
Models can propose plans, interpretations, challenges and drafts. A human decides what to accept, qualify, defer, reject or carry forward. That decision is recorded, but human approval does not make weak evidence stronger.
For brainstorming, casual research or low-consequence drafting, the extra evidence discipline may be more overhead than value. It is aimed at work where someone may later ask: what supported this, what was uncertain, and who decided?
The evidence-quality series covers citation support, amplification vs corroboration, stale evidence, human review, model-as-judge failure, unresolved disagreement and where scarce attention should go — the same failure modes the architecture is designed to preserve rather than hide.
Bring us a real workflow. Tell us what decisions are being made, what method already exists, where AI enters and what must remain controlled. We will respond within a week. Please do not include confidential evidence in the form.