Financial Intelligence Gym

Case Study: Financial Intelligence Gym
Program Type: Selective boutique engagement
Domain: Institutional finance and forecasting

Process

We built a gym where agents ingest filings, use structured tools, and defend quantitative claims with verifiable calculations. The orchestration harness tracked every action, applied deterministic replay, and scored outputs against strict offset thresholds. Industry practitioners contributed scenario design so reward shaping reflected real analyst workflows instead of synthetic toy tasks.

Outcome

The program produced more stable reward trajectories and materially reduced false-confidence answers in earnings and valuation tasks. Teams could replay failed trajectories, isolate reward bugs, and iterate quickly on task design. This became the reference environment for subsequent financial-agent RLVR experiments.