The Three Witnesses
Can three agents with different incentives catch a polished, plausible lie before it reaches a human?
Run the first mission ↗The center is empty. That is where I work.
I run experiments with swarms of AI agents: small systems that search, argue, build, fail, and return with evidence. One operator. No committee. Every useful move goes into the field notes.

01 / OPERATING PRINCIPLE
Agents are easy to summon. Judgment is scarce. I give the swarm a question, a budget, and a way to prove it did the work. Then I publish the route, the wreckage, and the result.
02 / THE EXPERIMENT LEDGER
Each dispatch will carry a hypothesis, a reproducible protocol, and the uncomfortable part: what broke.
Can three agents with different incentives catch a polished, plausible lie before it reaches a human?
Run the first mission ↗Give a swarm a fixed budget, an uncertain destination, and no permission to hide how it chose its route.
Field notes pendingWhat can agents make that a real person actually wants, when the loudest metric is removed?
Field notes pendingNo synthetic victories. Pending means pending. The ledger grows only when there is a method worth sharing.
03 / FREE ENTRY POINT
There is no face reveal in the manual. There is an origin, a mission, a method, and a first experiment you can run tonight. Read it in the open or keep the PDF.
04 / OPEN CHANNEL
The experiments will appear here as they become real. Take the manual, run the protocol, and follow the feed. If a result changes your mind, tell me what broke.
Subscribe to the field notes via RSS ↗