Epistamate is built around a separation that becomes more important as models improve: model reasoning, governed evidence state and professional judgement are not the same thing.
Use the strongest available intelligence for cognition. Keep source identity, representation, model-facing attention, human authority and later change explicit in application-owned state.
These are controlled product observations, not client outcomes or comparative accuracy claims.
In one document-led run, 10 citation routes were reachable and 17 failed or were blocked. Epistamate kept route reachability separate from citation identity, claim support and substantive validity, and later analysed 11 unique preserved source contents after de-duplication.
A controlled first pass planned 14 targets, acquired 12 and left 2 explicitly unresolved or deferred. The synthesis still retained 11 knowledge gaps and 5 validation leads rather than presenting the first pass as complete.
In a synthetic specialist-advisory experiment, accepted evidence propagated through working propositions, diagnostics, findings, recommendations and affected deliverables in causal order while the consultant retained authority at each transition.
Consequential propositions remain durable objects that can be supported, qualified, contradicted, reviewed and reused independently of the prose around them.
Directness, source role, exact support, independence, applicability, contradiction, recency and sufficiency remain inspectable dimensions.
Epistamate does not calculate a universal confidence score or source-independence score. Evidence posture remains multidimensional and inspectable rather than collapsed into a single number.
Models can surface counterevidence and alternative explanations, but challenge changes canonical posture only when the evidence justifies it.
Missing evidence is a durable research object and possible Research Further target, not a disclaimer that disappears after drafting.
Inherited evidence can be reconsidered for freshness, applicability and contradiction. Inherited answer authority is not automatically reusable.
A research question can drive acquisition; a supplied report can drive verification. Both enter the same evidence and review spine.
Question → planning → acquisition → evidence assessment → review → synthesis → Research Further.
Workflow propositions → operational evidence → research-informed challenge → findings → strengthening action → reassessment.
Your method → material claims → mixed evidence → assessment → recommendation dependencies → controlled output → update / reassessment.
The target reusable object is closer to claim + exact evidence + provenance + applicability + independence + contradiction state + review history + version lineage. Same-topic continuity exists today; durable cross-topic canonical Knowledge remains under rebuild.
The Version 1.0 technical preprint records the current Research Workspace architecture and exercised case boundaries. Earlier DCBR and RegWatch papers document the architectural lineage and regulatory extension.
Architecture and case studies from the Epistamate Research Workspace, including bounded attention, authority separation, iteration and citation-support closure.
Canonical record for the original published architecture.
IP.com changed its public retrieval surface. Use its prior-art search and the publication identifier shown here rather than the old fragile deep link.
Published paper on domain-configurable bidirectional reasoning for regulatory intelligence and epistemic infrastructure.
Where Epistamate sits relative to frontier research agents, scholarly tools, RAG, knowledge graphs and AI governance.