Built around a research loop, enforced through AI.
Find a question, form an understanding, write it as a strategy, watch the result, let the feedback improve the next round. AI makes recording, testing and updating part of the work rather than a good intention.
What are the layers?
Taste
Which changes deserve attention, and which questions deserve research budget.
Harness
The data, tools and working mechanisms that organize research. People and AI both go through it.
Strategy
Understanding written as executable rules: read information, form a prediction, decide what to do.
Prediction
A testable judgment made at a point in time. It cannot be rewritten after the outcome arrives.
Everything is a learning problem.
Not just the strategy. The evaluator, the generator, the taste and the harness each have an objective, data, feedback and a rule for how they update.
Hand-written rules → learned scoring → learned generation → closed loop.
What we believe.
- P-01
Research can recurse. Markets won't restart for you.
History can be recomputed. Evidence can't be copied.
- P-02
Put numbers on uncertainty, then let reality grade them.
A probability per scenario, filed before the outcome.
- P-03
Design for the models of six months from now.
Scaffolding should be easy to delete.
Different disciplines, one loop.
Make the loop improve itself
Where self-improvement and RL apply in research, and how a person's taste becomes a model.
Build the workbench
Clear modules and contracts. Data, tools and evaluators that can be versioned and judged.
Bring the real problems
Say what you need. Write your rules down, run them, iterate on results.
Understand before you allocate
Every strategy comes with its basis, conditions and risks.
Building in public.
Come argue with us.
We share research, runnable versions, results and open problems. Stress-test them.