Source lineage
Apparent consensus can be one original result amplified through many papers, reports and commentary. Counting sources is not enough; independence matters.
A policy brief can contain real citations and still overstate what the evidence supports. A single study can be repeated through many secondary sources. Evidence from one population can be silently generalised to another. A disputed claim can disappear into smooth synthesis.
Epistamate is designed to preserve the support, scope, contradictions and gaps underneath policy conclusions — together with the review history.
Apparent consensus can be one original result amplified through many papers, reports and commentary. Counting sources is not enough; independence matters.
Evidence can be strong and still fail to transfer to the jurisdiction, population, mechanism or implementation context of the policy question.
Disagreement is often professionally inconvenient, so synthesis resolves it too early. Epistamate treats contradiction and qualification as information rather than defects.
What the evidence does not establish may be more consequential than what it does. Gaps should survive the report and become future research targets.
A future reviewer should be able to distinguish AI-proposed priorities from claims an analyst actually inspected and judged.
Prior research should be reusable with provenance and revalidation rather than rebuilt from scratch or copied forward without checking recency and applicability.
Consequential propositions remain linked to source identity and the material that supports, qualifies or contradicts them. A bibliography is part of the output; the support chain is the more important object underneath it.
Unresolved needs remain visible and can drive Research Further. The system is allowed to say “the evidence is insufficient” rather than manufacture a complete-looking brief.
Frontier models can improve acquisition planning, review prioritisation and adversarial challenge. Their suggestions remain advisory; model agreement does not count as independent evidence.
The long-term Knowledge layer is intended to carry prior evidence, gaps and review history forward while checking currency, applicability and contradiction for each new policy question.
Researcher decisions are explicit events tied to the research state. A senior reviewer can reconstruct what evidence was available and what the analyst actually did, without treating approval as evidence strength.
Official instruments, implementation evidence, expert analysis and commentary can play different roles. The engine architecture supports domain policy packs rather than relying on prompt instructions alone.
A policy claim can be normatively attractive yet weakly evidenced on implementation feasibility or stakeholder effects. Those dimensions should remain separable.
Assumptions and transferability conditions should remain visible so a reviewer can challenge them instead of inheriting them from the synthesis.
We are looking for consequential policy workflows where evidence traceability, gaps and reviewer provenance matter.
Start a conversation →