Minerval

Minerval is an open source project to build a database of claims about the world and the questions they answer, figure out what the important ones are, and ask LLMs to assess their epistemic statuses.¹ Its claim pages (example) are intended to be the place that one links to or consults for the canonical, neutral epistemic status of any claim, much as Wikipedia is the canonical (and at least aspirationally neutral) place one goes for a standard introduction to any given topic.

What to link to to prove your point on the internet is an underappreciated unsolved problem, even when you are 'demonstrably' right. None of the existing solutions are really adequate. Some random paper on Arxiv whose jargony abstract plausibly says something adjacent to the thing you're saying? A blog post by some guy your interlocutor doesn't know and doesn't trust? A news article? The Wikipedia page for an adjacent topic? A link to your conversation with Claude where you controlled the framing of the question? None of these things is likely to satisfy the objector.

Minerval decomposes every claim to its bedrock and every question into what it presupposes, weighs both against the evidence, and keeps the verdict current as the world changes. The graph is maintained by LLM administrators operating under a public constitution; every judgment carries a reasoning trace, and every decision is open to challenge. Like Wikipedia, the graph is a public good, and the payoff is what gets built on it: the site at minerval.ai, a browser extension that annotates the web by verdict, and an API and MCP server that ground AI agents in claims that have already been weighed.

If done properly, with today's best LLMs,² this should allow for actually good mass fact-checking, in the same way that Pangram allows actually good mass slop-detection.³

I am aware of the long history of failed attempts at similar projects, and why most ideas in the category are indeed doomed. I am also very aware of the challenges of doing actually good epistemic work across domains with even the best of today's LLMs. For more details on how I approach the problem, please read our constitution.

Where it's going

In the long term, I aim to turn Minerval into a central institution that organizes the funding of research and knowledge production across domains. To efficiently allocate attention in a scientific domain, we must first have a map of the work to be done, and then a mechanism for pricing that work.⁴ Minerval does both. By delegating funding decisions to grantmaker agents (docs/allocation.md), it aims to solve the information asymmetry problem in the market for science.

Status

I have built an eval suite measuring properties which, if robustly satisfied at the graph level by the multi-agent system that administers the graph, would ensure that it can scale efficiently without losing coherence or accuracy (stability, path independence, consistency, and adversarial robustness).⁵ Our current mapped subdomains were built for the Future of Life Foundation's Epistack competition, where Minerval was selected as a winner. They look about right, but we need to run larger multi-agent tests to be confident that these properties hold.

Who's behind this

Minerval is built by Jackson Hurley.⁶ At this time, Minerval is just me and Claude Code. I am actively looking for cofounders. If you like the vision and want to work on it, please reach out. If you know the perfect person, please put us in touch!

This repository

This repository is the whole system: the API and agent pipeline, the web app, the browser extension, the Claude Code plugin, the Lean checker, the evaluation harness, and the infrastructure. The rest of this README is for people who want to read or run the code. The documentation on the site covers the same ground for readers, including every agent's complete system prompt.

Design commitments

Most epistemic tools work at the level of documents: an article gets fact-checked, a page gets written up encyclopedically. But disputes live at the level of claims, and the same claim recurs across thousands of documents. Minerval takes the claim as its atomic unit and does the expensive work once. A claim is extracted, canonicalized, decomposed, and assessed a single time, then reused everywhere it appears.

The constitution sets out how that works. Claims decompose until they reach what no informed person actually disputes (§6). Two formulations are the same claim when they turn on the same considerations, and a claim and its denial are one node (§2, §3). Independent lines of reasoning are recorded as arguments, whose soundness is itself a claim (§7). Every assessment is one of six graded statuses with its reasoning attached, so a disagreement that can't be settled is mapped rather than forced to a verdict (§10). A status describes the evidence, not the debate: how actively a claim is disputed is recorded separately and shown beside it.

How it works

   SOURCE              INGESTION                     GRAPH
 ┌─────────┐   ┌──────────────────────────┐   ┌────────────┐
 │ URL or  │──▶│ Extractor → Matcher →    │──▶│ Postgres   │
 │ document│   │ Claim Steward            │   │ + pgvector │
 └─────────┘   └──────────────────────────┘   └─────┬──────┘
                                                    │ read
   GOVERNANCE (ongoing)                             ▼
 ┌─────────────────────────────────────┐      ┌───────────┐     web ·
 │ Claim Steward · Curator ·           │◀────▶│    API    │──▶  extension ·
 │ Consistency Checker · Contribution  │      │ (Fastify) │     MCP clients
 │ Reviewer · Dispute Arbitrator ·     │      └───────────┘
 │ Audit Agent · Grantmaker · Lookout  │
 └─────────────────────────────────────┘

Ingestion is the expensive, write-side work. Serving is cheap, with no LLM in the read path. The graph is maintained by ten LLM administrators, each bound by the constitution and each with a bounded domain.

Building the graph

• Extractor reads a source and surfaces the discrete, reusable claims it asserts. It is deliberately selective: the claims a reader would want checked, not every sentence.
• Matcher is the identity gate. For each proposed claim it searches the graph under multiple framings, including the negation, and decides whether to match or create. The other agents also call it before creating anything.

Keeping it honest

• Claim Steward owns a single claim end to end: it decomposes it, maintains its canonical form and arguments, sets its importance, and assesses it, re-judging as evidence and the claims it depends on change. Decomposing and assessing are one open-ended judgment, so they belong to one owner.
• Curator owns the connective tissue between claims: merging duplicates the Matcher missed, splitting conflations, proposing cross-claim edges. It never overrides a Steward's verdict.
• Consistency Checker reads a region of the graph at a time for verdicts that conflict with their neighbors' or don't follow from what they rest on, and raises each find with the Steward whose claim looks wrong. It writes no verdict itself.
• Contribution Reviewer, Dispute Arbitrator, and Audit Agent run governance: policy review of incoming contributions, adjudication of escalations and appeals, and sampled quality control over the system's own decisions.
• Grantmaker designs and stewards a funded mandate: it surveys the territory, values the work on offer, grows its own plan, moves budget between peer mandates, and may refuse money that would warp the graph.
• Lookout is a mandate's standing watch. Woken by a heartbeat or a trigger (the daily retraction poll, a poke), it reads the graph, the retraction record, and the open web, and raises candidates for its Grantmaker: a claim to reassess, a source to ingest. It judges relevance, never truth, and can neither write an assessment nor move money.

Instruments. A few more agents serve the administrators without administering anything. The Researcher runs one bounded investigation for the agent that launched it (replicate a finding, trace a statistic to its origin, map a literature) and reports back. The Solver makes one bounded attempt on a formal mathematical statement, with Lean and a computer-algebra sandbox, and reports to the claim's Steward. The Tagger labels claims with topic tags for navigation. The Extension Agent sits behind the browser extension, judging on-page phrasings against the graph and powering its chat; it never writes to the graph.

Model choice follows the value of the judgment. Narrow, saturating calls (matching, tagging, the Lookout's relevance call, the Consistency Checker's sweep) run on a cheap model via OpenRouter. The load-bearing epistemic work (extraction, stewardship, curation, arbitration, audit, grantmaking, solving) runs on the strongest available Claude model. Defaults live in src/llm/models.ts and src/config.ts, and every agent can be pointed at another model with its *_MODEL environment variable.

Allocation

One engine decides where the system spends its attention. Every potential action (assess this claim, ingest that source, plan that mandate) is a row on a shared action ledger. Mandates and people place money on those rows, and an action runs exactly when its allocations cover its expected cost; nothing else decides what runs. Money is denominated in owls, one owl per dollar of metered model cost. Each claim carries an importance score (roughly consequence-if-wrong times how actively it is disputed or consulted) that anchors what work on it is worth, so the most consequential claims are assessed first while minor ones stay searchable stubs until someone funds them. Mandates are public, funded programs of work on the graph, each stewarded by its own Grantmaker and watched by the Lookouts it posts. See docs/allocation.md.

Further reading

• docs/architecture.md: domain model, assessment semantics, workers and failure handling, persistence, serving surfaces.
• docs/policies.md: the operating policies the agents apply on top of the constitution.
• docs/mathematics.md: mathematics as the flagship domain (formal statements, the Lean checker, the solver, prizes).
• docs/mcp.md: the MCP server and its tools.

Surfaces

• Web app. minerval.ai, a Next.js app for browsing claims, decomposition trees, arguments, assessments, and contribution history, and for browsing and funding mandates.
• API. Fastify at api.claimgraph.io. Reads are public; anything that writes or spends model tokens requires a key. Interactive OpenAPI docs are at /docs on the API host.
• Browser extension. Reads the page with you, underlining each recognized claim by what the graph knows about it, with a chat grounded in the graph. Available on the Chrome Web Store; built with WXT in extension/.
• MCP server. Remote MCP over streamable HTTP at POST /mcp, with OAuth 2.1 so hosted clients (such as Claude.ai) can connect. Tools for searching and reading the graph, running the pipeline's judgments, and contributing. See docs/mcp.md.
• Claude Code plugin. Packages the MCP server with slash commands, a fact-checker subagent, and a skill. This repository doubles as its marketplace; see plugin/.

Repository layout

Path	Contents
src/	The API (Fastify), the agent pipeline (llm/, workers/), services, the MCP server, and the Drizzle schema (db/)
web/	The Next.js web app deployed at minerval.ai
extension/	The browser extension (WXT)
plugin/	The Claude Code / Cowork plugin
lean-checker/	The Lean checker service for formal mathematical statements
skills/	Skills the agents load for particular domains
corpus/	The evaluation harness: pinned document clusters, scoring rubric, LLM-judge scoring
tests/	Unit and database tests
scripts/	Operational scripts: seeding, backfills, repairs, eval content
docs/	Architecture, policies, allocation, mathematics, MCP, accounts, reputation, infrastructure

Running locally

The whole pipeline runs on a laptop: docker-compose provides Postgres with pgvector, and the job queue runs in memory with handlers identical to the SQS ones used in production. You need Node.js 22, Docker, and API keys for Anthropic (the strong-tier agents), OpenAI (embeddings), and OpenRouter (the cheap-tier agents). To run without OpenRouter, point MATCHER_MODEL, TAGGER_MODEL, LOOKOUT_MODEL, and CONSISTENCY_MODEL at a Claude model instead; .env.example documents each one.

docker compose up -d          # Postgres + pgvector
cp .env.example .env          # fill in the API keys
npm install
npm run db:migrate
npm run dev                   # API + workers on :3000

npm test runs the unit tests; npm run typecheck checks types.

Evaluation

Agent changes are graded, not eyeballed. The corpus harness runs the real application over pinned clusters of source documents, drains the pipeline to quiescence, and has an LLM judge (deliberately a different model from the agent under test) score the resulting graph against the constitution. The graph-level evals for stability, path independence, consistency, and adversarial robustness build on it. Results, costs, and recorded replays are published at minerval.ai/docs/evals; the harness is in corpus/.

Contributing

To the knowledge graph: submit challenges, evidence, and proposals through the web app, the API, or the MCP server. They flow through the reviewed governance pipeline described above.

To the code: issues and pull requests are welcome. Start with docs/architecture.md to get oriented.

License

The code is licensed under the Apache License 2.0; copyright Minerval, Inc. See NOTICE.

The claim graph content (claims, assessments, arguments, and the nanopublication exports built from them) is dedicated to the public domain under CC0 1.0. Each nanopub records the dedication in its publication-info graph via dct:license. Contributions to the knowledge graph are accepted under the same CC0 dedication.

---

"The owl of Minerva spreads its wings only with the falling of the dusk." (Hegel)

Understanding, Hegel thought, arrives only in retrospect. Minerval is an attempt to do better: to map claims as they are made, not after the dust has settled.

Footnotes

1. Claims are canonical forms of public claims about the world. More details in part II of our constitution. ↩

2. Nobody will use this, and no one should use it, if the outputs are noticeably worse than what they could get by opening a new tab and simply asking their own frontier model. ↩

3. Including via our browser extension or, ideally, since no one uses browser extensions, direct integration with platforms such as LessWrong, Substack (similar to what they've done with Pangram), or X. ↩

4. A complex problem. See Allocating Attention in Claimspace. ↩

5. If Minerval succeeds at becoming a canonical source, people will try to influence it. Minerval invites contributions and new evidence, but that comes with substantial attack surface area for bad-faith actors. The plan here is to set up a test environment and give agents the task of manipulating the graph's administration into supporting their position, then keep adjusting until relentless frontier agents robustly fail at red-teaming. ↩

6. I have incorporated Minerval, Inc. as a vanilla Delaware C-Corp, but I have no immediate plans to either raise money from VCs or provide excludable goods and services, with the very minor exception of the chat feature in the browser extension/API.