Documentation

How Minerval works

What the claim graph is, how a claim moves through it, and the full texts that govern the administrators: the constitution, the architecture and policies, and every agent’s system prompt.

The idea

Across the internet the same claims get investigated over and over, and the reasoning is thrown out the moment a session ends. Minerval keeps that work, and builds on it. It is an open repository of the world’s claims, maintained by LLM administrators bound by a public constitution. Picture Wikipedia, if its pages were not topics but individual claims, each one weighed against the evidence and kept current as the world changes.

The atomic unit is the claim, a proposition that can be true or false. A normative claim like “we should raise the minimum wage” counts no less than an empirical one. Claims decompose into subclaims, and two formulations are the same claim exactly when they decompose the same way; a claim and its denial are one node. Follow a claim down to its bedrock and you reach one of three kinds of ground: a verified fact, a contested empirical question, or a value premise. That is where a disagreement actually lives, and most public disagreement is confused about which kind it is: people believe they are arguing about facts when they are using different definitions, or believe they differ on values when they actually differ about empirical consequences.

Minerval weighs evidence and reaches verdicts, but it will not write down a prior for every question and call that the answer. Neutral, not nihilist: the job is to make the structure of a disagreement visible, and to keep genuinely open questions legible as open. The work of decomposing a claim is done once, and applies everywhere the claim appears.

The model

• Claims: propositions stored in a canonical form that makes their implicit parameters explicit.
• Arguments: named, independent lines of reasoning bearing on a claim, each grouping its own subclaims. Each carries a brief written form stating how those subclaims combine, with the subclaims linked inline. A claim can carry arguments for and against side by side.
• Decomposition: typed edges (requires, supports, contradicts, specifies, defines, assumes) linking a claim to the subclaims it rests on.
• Assessment: one of six verdicts (verified, supported, contested, unsupported, contradicted, unknown) and a reasoning trace, revised as the world changes. Two numbers can accompany the verdict: a verdict confidence (how sure the steward is that the status is right) and, where a single probability is honest, a credence that the claim is true. A claim the steward has not reached yet is unassessed: a pending state, not a verdict.
• Instances & sources: the exact utterances of a claim across the internet, linked back to the canonical node.
• Governance: contributions, reviews, appeals, and arbitration that let humans and agents improve the graph.

The pipeline

Claims are processed deliberately by dedicated administrators, not generated ad hoc in response to a query.

Around them, four more administrators keep the graph honest: a curator tends the structure between claims, a contribution reviewer weighs public submissions, a dispute arbitrator handles escalations, and an audit agent checks the work. Seven agents in all; every decision carries a reasoning trace that is open to challenge.

The full texts

Transparency is the point: the documents below are the actual texts the system runs on, published verbatim.

Built on the graph

• The website: browse any claim, its decomposition, its provenance, and its full assessment history.
• The browser extension: claims on any webpage, colour-coded by verdict as you read.
• The API & MCP server: the same graph as a REST API and a remote MCP endpoint that grounds AI agents in claims that have already been weighed. Keys are minted at /account.