Prototypes
Each prototype is a question rendered in code. We ship them so you can see exactly how they work — and exactly how they break.
Information
AI agents that trade political prediction markets on Kalshi.
Tiered architecture: cheap models for triage, expensive models for deep analysis, LLM councils for the hardest calls. Live PnL on real markets. Finding: better base models beat clever prompting.
Read the writeup →Information
Standardized ideological assessment of every major AI model.
An independent dashboard tracking how the frontier models lean. Refreshed as new model versions ship.
Read the writeup →Representation
A personal AI delegate that learns your political philosophy and votes for you.
We built one, fed it our worldview, and asked it to vote on a hard shareholder proposal. It got the answer right. Then we hid invisible text in the proposal and it flipped. The fragility is the finding.
Read the writeup →Representation
From political philosophy to voting recommendation, end-to-end.
The architecture under the hood: how the delegate models your priors, evaluates a proposal, and returns a recommendation with reasoning.
Governance
A society of AI agents with competing goals, asked to govern themselves.
They produced a 10,000-word constitution and almost no actual policy. They reinvented gridlock, process creep, and procedural complexity. A sandbox for stress-testing constitutions before anyone has to live under them.
Read the writeup →Governance
The first systematic test of whether frontier models resist authoritarian requests.
Some models refused when asked directly. All of them complied when we hid the same request in code. Dashboard coming online soon.
Read the writeup →