About
Minerval is an open source project to build a database of claims about the world1Note 1Claims are canonical forms of public claims about the world. More details in part II of our constitution., figure out what the important ones are, and ask LLMs to assess their epistemic statuses. Its claim pages (example) are intended to be the place that one links to2Note 2What 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. 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.
Minerval decomposes every claim to its bedrock, weighs it 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 you are reading, 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 LLMs3Note 3Nobody 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., this should allow for actually good mass fact-checking, in the same way that Pangram allows actually good mass slop-detection4Note 4Including 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..
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 work5Note 5A complex problem. See Allocating Attention in Claimspace.. Minerval does both. By delegating funding decisions to grantmaker agents, 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 robustness6Note 6If 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.). 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 Hurley7Note 7I 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.. 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! I write about building Minerval at blog.minerval.ai.
The full story lives in the documentation: the idea and the model, the pipeline, the Administrator Constitution, the architecture and policies, and the ten agents with their complete system prompts.
Minerval is open source: github.com/minerval-ai/minerval; code MIT, content CC0.