Train
Training data
Worked examples and training records built from real financial decisions, reviewed by practitioners.
See training dataDissei Data is an AI-native finance firm. We build safe, reliable frontier environments, evaluations and data, using our unique aligned vertical structure.
Looking for financing or a lending partnership? Visit Dissei Lending.
Latest research
- ResearchMore thinking made Gemini 3.8 Flash less likely to finishIn our financial-judgment environment, Gemini 3.8 Flash at low reasoning effort submitted an answer in all 21 trials; at high effort, the setting behind our earlier results, it submitted in 7. Medium, its default, fell in between. This is one case, and our repeat was too small to confirm it, but the gap is larger than anything our fixes to the agent produced.
- ResearchOnly if you ask: planting wrong numbers in a 20-step credit worksheet with three modelsAccurate calculation from scratch does not guarantee reliable handling of an inherited error, even when the model can correct that error in an audit.
- ResearchSaying 70, meaning 40: guessing probabilities with five mid-tier modelsWe asked five mid-tier models for probabilities on 3,010 resolved stock-price questions, each asked twice, on two scales. Six of the ten sets of stated probabilities told us no more than a coin would.
What we doReal decisions,
Real decisions,
rebuilt as environments.
Evaluate
Sealed evaluations and graders
Assess financial judgment against fixed criteria, with evidence behind each score.
See sealed evaluationsRun
Environments
One real transaction, rebuilt as it stood on a fixed date. Evidence stops at the decision date, the agent works with permitted tools, and its decision is graded against criteria fixed before any model runs.
See environmentsSource
From our own lending and public records.
Each case starts as a dated record of a real decision and the outcome that followed. We own all of this data. In cases from our own lending, we change identifying detail and keep the reasoning.
Read how we build casesWho we work with
AI labs
Research and data for training and evaluation
Off the shelf from our existing cases, or built to your specification.
For AI labsEnterprises
Evaluations for your finance models
Evaluate finance models, including your own, or build on our environments and data.
For enterprisesFinancial judgment leaderboardFinancial judgment,
Financial judgment,
measured.
See how AI models reason from financial evidence. We compare them across seven reasoning categories and score each answer on a 0–100 continuous reward scale, not accuracy.
Scores by model, across all evaluations.
Published release
GPT-5.6 Sol: Observed score / 100 45.53. Claude Opus 4.8: Observed score / 100 44.26. Muse Spark 1.2: Observed score / 100 42.41. Kimi K3: Observed score / 100 39.76. DeepSeek v4-flash: Observed score / 100 25.49.
Observed score / 100
What do you sell?
We sell five products: training data, agent trajectories, sealed evaluations and graders, long horizon workflows and environments. Each comes off the shelf from our existing cases, or built to your specification.
Environments run inside your own perimeter. If you already run your own environment, we deliver the underlying data, tasks and reward components instead.
Why your data?
It records real decisions with known outcomes, not scraped text or opinion written to order. Each case shows what was known on the decision date, what was decided and what happened next.
We put our own capital behind every deal that we fund, so we carry the cost when our judgment is wrong. The firm that makes the decision lives with the outcome.
Where does the data come from, and do you have the right to train on it?
Our data comes from our own lending and from public records. We own all of it.
Each asset we deliver says where its right to train comes from. It also carries a contamination statement. Each product agreement states the access and use rights.
Tell us what your model
needs to learn.
Each product agreement states its access and use rights. Your training never touches our infrastructure.