The research goal: can AI output be stored and used as probability instead of fact? Every model produces probabilistic output, and most of the industry renders it as a single confident answer. That is where the damage comes from. This project tests the alternative: keep the distribution, keep the confidence, and score every claim against what actually happened.
Financial markets are the testbed because they grade the homework. A prediction about a price or a filing resolves on a known date, so the system's beliefs can be checked against reality at scale. It is internal research, not a product and not a revenue stream, and its findings feed every other project in the lab.
A six layer system: a dashboard on top, an orchestration gateway underneath, then control, data, and intelligence layers over a memory and validation store. Market data, SEC filings, and news feed a set of specialist AI agents for news, technical, and fundamental analysis, with an orchestrator combining their output.
The memory layer is the experiment. A belief validator checks new claims against stored ones, a trust labeler tracks which sources and agents have been right before, and a performance tracker scores predictions against outcomes, so confidence comes from a record rather than from the model's tone.
What the lab built
At a glance
Market data, filings, news
Specialist agents form claims
Claims kept as probabilities, not facts
Outcomes grade every prediction
Trust follows the record
