Context
Most trading systems — human or machine — fail the same way: unchallenged conviction. One model forms one opinion and acts on it. My doctoral research asks what changes when action has to be earned through argument: multiple AI agents with different worldviews debating every position before capital moves.
The problem
A single model can be confident and wrong, and nothing in the loop will say so. Quantitative failures rarely come from weak signals; they come from strong signals nobody stress-tested. The interesting question is not whether AI can trade. It is whether AI can doubt.
Shape of the system
Several agents with deliberately different priors argue each thesis. They cross-examine one another, and a gating mechanism scores the quality of that disagreement — not whether the agents agree, but whether the position survived real attack. Only earned consensus reaches the market. The specifics of the gate are the research.
Where it stands
Backtested across a full year of market data, where it caught trades that felt right and weren't. The first paper from this work was accepted at an international conference in 2026. The thesis continues.
In practice
- Agents designed around opposing priors
- Cross-examination as a first-class mechanism
- A gate that measures the quality of disagreement
- Execution only through earned consensus
- Active doctoral research
Instruments
Python · LLM APIs · Custom deliberation protocol · Backtesting framework