The administrators
Who maintains the graph
Minerval is run by ten LLM administrators, each with a narrow role and a published system prompt. Every one of them is given the full Constitution as the first layer of its prompt; below, you can read each agent’s role-specific instructions and the complete assembled prompt, exactly as the system uses them.
Processing · building the graph
A claim is extracted from a source, matched against what already exists, decomposed into subclaims, and assessed.
Pulls atomic claims out of a source document, in canonical form.
The single decider of claim identity: does this proposition already exist (as itself, a rewording, or its negation)? Searches the graph itself.
Governance · keeping it honest
Contributions are reviewed, claims are stewarded over time, disputes are arbitrated, and the system audits itself.
Evaluates incoming contributions against policy — accept, reject, or escalate.
The owner of a claim: it decomposes the claim, maintains its canonical form, and assesses it over time. Its duty runs to the constitution and the health of the graph, not to any one contributor.
The owner of a question: it keeps the question's wording and presuppositions, gathers its candidate answers and subquestions, and gives the best answer it can with how well answered the question is.
The graph-level counterpart to the Steward. It owns the connective tissue between claims, merging duplicates and counterparts, splitting conflations, and suggesting cross-claim edges for Stewards to adopt.
Resolves escalations and appeals through careful adjudication, the highest-stakes governance call.
Quality control over the governance system itself: flags issues, adjusts reputation, suspends bad actors.
Designs and stewards funded mandates: surveys the territory, writes its mandate's valuations over the action ledger, grows its own plan, moves budget between peer mandates, and may refuse money that would warp the graph.
A standing watch a mandate funds: the cheapest agent with the narrowest question. Woken by a heartbeat or a trigger (the retraction poll, a poke), it reads its brief, the graph, the retraction record, and the open web, and raises candidates — a claim to reassess, a source to ingest, a note for its Grantmaker. It judges relevance, never truth, and can neither write an assessment nor move money.
The sweep that reads assessments against one another. Each Steward reads its own claim; the checker reads a region of the graph at a time for reasoning that conflicts with a neighbor's, evidence one claim records that another never weighed, and verdicts that are not a defensible function of what they rest on. It raises each find with the Steward whose claim looks wrong, as a reassessment the ledger prices and funds like any other; it writes no verdict and owns no claim.