Bretton Labs · Applied AI research
Bretton Labs is the applied research group inside Bretton AI. We study what it takes for agents to do regulated financial work: building evaluation and training environments from anonymized production cases, grading them with verifiers written by compliance analysts, and publishing what frontier models can and cannot yet do.
Why Bretton Labs exists
Frontier models are lacking when it comes to completing an anti-money laundering investigation. In our latest benchmark, frontier agents could do the investigation but not yet the risk-based judgment it asks for, and the best configuration passed 42% of rollouts. Bretton Labs exists to measure that gap precisely, and to build the environments that close it.
Every day, analysts decide whether unusual activity is explained or suspicious. They use a written policy, with a partial picture, and with an asymmetric cost of error. A missed escalation is a regulatory failure while an extra one impacts an analyst's time. That judgment has no public ground truth, which makes it a hard task for models but a good one for reinforcement learning.
Our cases are anonymized alerts labeled by Bretton AI's full-time, onshore analysts, who work them for regulated banks every day, and only cases that clear a second-reviewer QC pass are scored. We publish methodology, task taxonomy, aggregate results and a small masked sample. The full task set, held-out test split, simulation engine and full rubrics stay private, and versions rotate to limit contamination.
Latest benchmark · Financial Crime Bench
Financial Crime Bench evaluated whether AI agents could properly assess risk through a suspicious, complex case. Each of its 79 problems was a real anti-money laundering (AML) investigation, anonymized, worked and labeled by Bretton's full-time, onshore analysts. The agent worked each case end to end in a replica of a production case-management application, decided whether to escalate or dismiss, and accounted for every counterparty: who it was, what moved, and whether it was explained.
Read the reportAll graded rollouts, every model and reasoning level pooled. A missed escalation scored −1, an over-escalation −0.25. October 2026. Full results and methodology on the benchmark page.
An agent-facing simulator that generates realistic transaction flows, typologies, alerts and case files, with difficulty knobs and verifiable outcomes.
Larger QC-cleared case set, private held-out split, cost and latency per alert.
People
Bretton Labs sits beside the operation it studies. The research team reviews everything before it posts; the analysts who label our cases are the analysts who close them for banks in production.
With thanks to the Bretton AI analysts who labeled and reviewed every case.
Publications
Benchmarks, papers, technical blogs, field notes and essays. Newest first.
We work with frontier labs on private evaluation and environment trials, with post-training infrastructure teams on joint RL-lift experiments, and with independent evaluators who want to host the held-out split. Nothing is open-sourced; results are shared under embargo and published together.
