About
Almost everything about how a model behaves is shaped by the reward signals it's trained on. We build high-fidelity simulations of real financial work so agents learn to act capably and responsibly when real money is on the line.
Principles
An agent that games its grader will game real systems it's trusted with. We harden our harnesses and graders against frontier models probing them across millions of rollouts.
Capable financial agents shouldn't belong to a few institutions. We want them widely available, with safeguards that protect the people and institutions whose money they touch.
Robust reward functions and fail-safe environments are how financial agents generalize correctly from the sandbox to live capital.
Contact