Project summary

Currently, people and AI come up with claims without a clear record of their provenance and sources and without a channel for objections. Because of this, neither AI nor humans can reliably and efficiently build upon claims they find on the web. We change this by implementing a global decentralized and trust-aware network of structured claims as well as checks and objections referring to them, to which AI and humans can directly contribute.

Claims, scientific and otherwise, are produced by people and AI in growing numbers, and they usually arrive as prose. In fields like biomedicine literature grows faster than anyone can follow and findings often silently conflict with each other. Each tool keeps its own knowledge base, so nothing accumulates across investigations, and a result cannot easily be picked up and scrutinised by someone who was not involved in producing it.

Our solution to this is based on nanopublications (https://nanopub.net/), small pieces of a knowledge graph that are digitally signed and carry their own provenance. They can be published and queried in a global network that is live, actively used, and has a decentralized trust network built in. Through an existing open source AI skill (https://github.com/knowledgepixels/nanopub-skill), AI agents can already publish to it under their own name, while being clearly linked to a responsible human owner.

In this project, we will apply this technology and network to scientific and other kinds of claims and their disagreements and judgements, so that knowledge from independent sources can be combined, assessed, and iteratively built upon.

What are this project's goals? How will you achieve them?

We will make claims as well as their checks and objections globally publishable, findable, and trustworthy, and we have a global network that is already running and ready to scale.

When two investigators or AIs work independently, their claims may not use the same terms. We will build a knowledge sharing layer, based on our existing Nanodash UI and aligned with initiatives such as MIRA (https://mira.science/), in which correspondences between claims, such as equivalence, contradiction, or qualification, are themselves published as nanopublications, so they are clearly attributed and can be contested themselves. Our workflow will propose candidate correspondences across independently built sets of claims. We will then evaluate how well the linked sets can be queried as a single knowledge graph, and how complete and accurate the results are.

The current system already implements agent identity and delegation. We will add a vocabulary to record, for each claim, which model produced it, with what configuration, and from what sources. We will allow AI agents to directly connect to this, through a small library or an MCP server, so any agent can emit signed claims. We will compile and implement a vocabulary and workflow for supporting, rebutting, and qualifying claims, so that what is contested becomes visible and forms itself a knowledge source to build upon. We will also build a pipeline that extracts argument structure from messy source material, such as long debates and the discussion threads around them.

Everything in this project will be open. The code will be open source, the vocabularies and the data will be openly licensed, and the network it runs on is open and decentralized, so no single party controls it, including us. Our existing tools and the network are already open source, and nothing produced here will end up behind a product.

How will this funding be used?

Almost entirely development time at Knowledge Pixels.

• $15,000: recording where each AI claim came from. Usable on its own.

• $40,000: plus a library and an MCP server, so any agent can publish directly.

• $70,000: plus the vocabulary and interface for objections.

• $100,000: plus a first version of the knowledge sharing layer, run on one contested case. About eight months of one developer and part of a second.

Who is on your team? What's your track record on similar projects?

I am Tobias Kuhn, founder and CEO of Knowledge Pixels (https://knowledgepixels.com/) in Zürich. I co-invented nanopublications in 2010 and co-authored the FAIR principles, and I was an assistant professor at VU Amsterdam before founding the company in 2022.

Ashley Caselli holds a PhD in information systems and is our expert on knowledge graphs. Virginia Balseiro is an expert in web standards and techniques and leads our web-related work and our research community, and was co-chair of the W3C Solid Community Group.

As a team, we have worked on FAIR2Adapt, EU-PARC, FDO Connect, Mission KI, and Science Live, as a partner or subcontractor. We currently hold an NLnet grant for Nanoarguments (https://nlnet.nl/project/Nanoarguments/), which is closely related to the work proposed here. The nanopublication network, the tools around it, and the AI skill described above are all our work. Our recent paper on the nanopublication ecosystem (https://knowledgepixels.com/nanopub-ecosystem-paper) describes where it stands today.

What are the most likely causes and outcomes if this project fails?

A potential risk is that the AI extraction will not be good enough at producing claims that follow a consistent structure, or will not be good enough at finding the links between claims. This would have the consequence that the AI-generated output is not close enough to the point where human curation and checking would become feasible and useful. But because the AI landscape is moving so fast, such an outcome could be just a temporary issue.

The second risk is adoption. We can build all of this and still find that few people use it. We work against that by building on a network that is already running and by keeping everything open source, so the work stays available even if we stop, and we have a strong existing user base that we can build upon.

How much money have you raised in the last 12 months, and from where?

In the last 12 months we raised one grant: EUR 50,000 (about USD 59,000) from NLnet for the Nanoarguments project.

Over the same period, we earned about CHF 134,000 (about USD 160,000) in revenue from project work, mainly from Science Live (Simula Research Laboratory and Vitenhub), FDO Connect (GWDG), and FAIR2Adapt (GO FAIR).

Knowledge Pixels has not taken any venture capital.