Agents will do the credit work. We measure where they fail.
Foundational research on long horizon agents, underwriting benchmarks and loan risk. This is separate from Intelligence, which publishes market data for lenders.
Paper · NeurIPS 2026
SWE-Marathon
We co-authored SWE-Marathon, a benchmark of software tasks that run for tens of millions of tokens. Over a long run, agents hack the reward and drift from the intent they were given. The same risk applies when an agent operates collateral. Our angle: live monitoring of what an agent does, tied to the assets it touches.
Visit SWE-Marathon →Trusted by
- xAI
- Moonshot AI
- Z.ai
- Alibaba Qwen
- Tencent
- Meta
- Mercor
Benchmark · In progress
UWBench v2
We are partnering with Tobias Pfuetze on version 2 of UWBench, the open benchmark for credit underwriting agents. It scores what the agent did, not what it said: the files it opened, the information it requested, and whether each source it cited exists.
Read about v1 →Model · In build
Foundation risk model
A foundation risk model for equipment loans. For each loan it gives the probability that the loan goes bad, and a confidence on that probability. A lender sees the risk and how sure the model is.
How it learns →Train to a date. Let the future happen. Measure the gap and close it.
A risk model is a bet on the future value of the collateral. The only fair test is time. So we freeze what the model knows at a cutoff date, let the market move, and score the model against what happened.
Step 1
Post-train to a cutoff
We post-train the model on data up to a fixed date. It sees nothing after that date.
Step 2
Predict
For each loan it gives the future value of the collateral, the probability the loan goes bad, and a confidence on both.
Step 3
Let it play out
Resale prices, rental rates and loan outcomes arrive. The predictions stay locked.
Step 4
Replay and measure the delta
We replay the model on the same loans and compare its forecast to the outcome. The delta shows its blind spots and its biases, for example a class it valued too high.
Step 5
Bridge the gap
We correct for the delta and post-train again at a new cutoff. Each cycle keeps the model's forward valuation close to what the market will pay.
Want to work on this, or fund it?