Core infrastructure
Public Knowledge Layer
A public research library that people can read and AI agents can use in further investigations. Findings link to their sources, show where evidence conflicts and retain a history of corrections.
Can a shared library improve the next investigation?
Researchers repeatedly assemble the same background: sources, definitions, competing explanations and unresolved questions. The Public Knowledge Layer is the shared research library we are building to preserve that work. Connecting findings across sources lets each investigation build on earlier work, rather than start with a new search or an empty conversation.
The AI-agent team's role is to use and extend the library through research with people. Reviewed findings can return to the shared collection; generated output does not become accepted knowledge automatically.
The comparison protocol assigns the same research task with and without access to a selected part of the library. Effort saved and quality gained are separate evaluation questions. The review criteria cover new findings, missing or unsupported claims, overlooked sources and outdated information; human review effort also counts.
Keep sources and uncertainty attached.
Every entry must explain a research finding, link to the source passages behind it, and show its date and remaining uncertainty. Corrections must show readers what changed.
This material is for people and AI agents beyond our team to inspect, question and build on.
- explanations that compare several sources rather than summarise one document
- established, contested, uncertain and unknown states
- dates for each claim and its last verification
- readable pages for people and machine-readable access for AI agents
Does shared context also spread shared mistakes?
Reviewed knowledge can still be wrong or become outdated. Agreement between AI agents using the same source collection is not independent confirmation: they may inherit the same blind spot.
The evaluation tests whether agents return to primary evidence when context is stale or disputed, and whether corrections change dependent conclusions. Reuse must preserve the ability to question what has already been accepted.
Preparing the first research collection.
We are selecting sources and test questions, beginning with Gonka's network economics and AI hardware research. Trials combining the library and AI-agent team are still ahead. The first public collection has a defined scope: one question, its sources and unresolved disagreements.
The library and agent access are not yet released. This website currently describes the projects and their status; its JSON index is not the knowledge library itself.
Have a source, contradiction or correction?