FJB: a dual-rubric benchmark for financial judgment. Read the working paper
Outcome-grounded RL environments and evaluation for financial judgment

Teaching machines financial judgment.

We take real deals, work them the way an analyst would, and grade the answer against what actually happened. Built for frontier labs, by practitioners.

The thesis

Quality cannot be crowdsourced.

The hard part of this domain is not volume. It is judgment, and judgment does not come off an assembly line.

An annotation farm can label a million examples. It cannot tell you whether reported EBITDA is real cash or accounting noise, where a first-lien claim sits in the recovery waterfall, or why a covenant decides the outcome. That takes practitioners in a room, building the framework for how to approach the problem before a single task is written.

This is why the work has to come from people who have been right and wrong with real money on the line. Quality data and quality benchmarks are downstream of quality judgment. Get the framework wrong and you have scaled the wrong answer. We do one thing: finance.

And we scale it the way no vendor can. Our originators underwrite real deals, and every deal becomes new ground to learn from.

The asymmetry
A model can be checked. A market never offered to.

Code got smart because every answer has a verifier: the tests pass or they do not. Finance never had that signal for the judgment call itself: nothing checked how a decision was reached against what later happened, so it had no automatic way to know when it was right.

The ceiling
So finance trained on opinion.

With no objective checker, the best you can learn from is consensus. And consensus caps you at average. You cannot beat the market by memorizing it.

The mechanism
Every deal is an environment. Its outcome is the reward.

A real deal already holds the missing signal: what actually happened. Frame it as a problem to solve, and the realized result becomes a reward a model can train against truth, not against opinion.

Products

Judgment, captured so a model can inherit it.

Nobody is born an analyst. Judgment accumulates, deal by deal, in people who have carried the risk. We capture that wordless calculus in forms a frontier lab can train on: artifacts you own that drop into your own RL and SFT pipelines, not a platform to adopt. Every product is finance, and only finance.

ValuationUnderwritingRiskStructuringMarkets
Reinforcement learning
Durable environments where real financial work happens: long horizons, primary documents, the tools an analyst actually uses, and an outcome that can be checked. This is reinforcement learning with verifiable rewards (RLVR), applied to finance. We preserve the full complexity of each deal, whether it is the first or the ten-thousandth.
Long-horizon deals
Underwriting that unfolds over days, through ambiguity, dead ends, and revision. We capture work that reflects how real diligence happens while preserving the signal needed to improve a model.
Off-the-shelf datasets Coming soon
Prebuilt finance corpora, curated for signal and reviewed by practitioners, structured to drop into your training stack without translation work.
Benchmarks and evals
Quality is easy to reduce to the wrong metric. FJB measures domain-faithful lift: judgment that holds against what actually happened, graded on a dual rubric across more than seventy competencies of financial judgment.
Analyst trajectories Coming soon
Full traces of how an expert works a deal: the documents pulled, the model built, the calls reversed, the thesis defended. Models learn the shape of the work, not just the finished answer.
Supervised fine-tuning Coming soon
Demonstrations that set the right prior, captured through tooling that lets analysts work naturally while preserving the operating judgment behind every decision.

FJB scores a model twice. Once against industry best practice, and once against how the deal was actually underwritten.

The distance between those two scores is the judgment gap. It is the headroom a model has left, a quantity no existing benchmark measures.

What FJB measures
01
Willing to pay
Who benefits from an outcome, and whether the incentives of owners, sponsors, and counterparties actually line up with it.
02
Value of parts
What the pieces are worth: the capital structure, the priority of each claim, and what the assets genuinely recover.
03
Future cash flow
Whether reported earnings convert into cash that can meet obligations, and how much runway there really is.
04
Known and unknown risks
The terms, exposures, and second-order risks that decide an outcome but rarely sit on the surface.
Reinforcement-learning environments

Every case becomes a training environment, not a static dataset.

A real financial situation is rebuilt into a set of graded tasks a model can act in and be scored on, with rewards that check against the record. How that rebuild happens stays in-house, but the controls that make it trainable do not: point-in-time splits so nothing leaks from the future, and every task adjudicated by practitioners before it counts. What reaches a model is the judgment of people who have actually carried the risk.

01
Verifiable reward
Each task is graded against evidence in the record, so the signal a model trains on is defensible rather than a matter of taste.
02
Point-in-time by construction
A model sees only what was knowable at the moment of decision. The outcome is held back, so there is no hindsight to reward.
03
Authored by practitioners
Environments are built and reviewed by analysts who have worked these situations, and admitted only once they clear that bar.
Seven modes of reasoning

Each task targets a capability frontier models still lack.

The categories isolate distinct kinds of judgment, drawn from how our analysts actually work a problem. The rubric in each category encodes that discipline.

Strategic
Read a situation and choose a course of action under uncertainty.
Diagnostic
Locate the cause beneath a set of symptoms in the record.
Quantitative
Derive and defend a number from the underlying figures.
Counterfactual
Reason about what a different choice would have produced.
Predictive
Forecast forward from only what was knowable at the anchor.
Explanatory
Account for why an outcome occurred, with evidence.
Comparative
Weigh two situations against each other and justify the distinction. Across all seven, the rubric rewards reasoning that holds up against what actually happened, not against consensus opinion.
For the team evaluating us

Questions a model team tends to ask first.

What is an environment, concretely?
A real deal, rebuilt as a set of graded tasks a model acts in and is scored on. Long horizons, primary documents, the tools an analyst uses, and an outcome that can be checked against the record.
Markets are noisy. How can an outcome be a reward?
The task is framed at the moment of decision and graded on the reasoning that was defensible then, not only on the dollars that followed. A sound call that lost money still scores. A lucky one does not.
How do you keep the answer from leaking into the eval?
Point-in-time by construction. A model sees only what was knowable at the anchor, the outcome is held back, and the framing is checked for the tells that quietly hand over an answer.
Why finance, and not code or math?
Code and math already have verifiers, so models compounded there. Finance's new benchmarks check answers; nothing yet verifies the judgment behind a decision against what actually happened. Closing that gap is the whole company.
Who actually builds these?
People who have carried the risk, not an annotation farm. Quality data is downstream of quality judgment, so the framework comes from practitioners before a single task is written.
Get in touch

Tell us what you are training.

Whether you are a frontier lab working with finance data or an institution with deals to put to work, we will scope it with you directly.

Mutual NDA before anything sensitive is shared
A real person replies, usually within two business days
What a first call covers
How an environment is specified, end to end
A rubric excerpt and how a task is scored
The point-in-time controls that keep an eval clean
Delivery formats and how they drop into your stack
Please enter your name.
Please enter your organization.
Please enter a valid work email.
A sentence or two on scope, scale, or timeline.

Message received.

Thank you. We will be in touch shortly to scope it with you.

Building the verifier finance still lacks.