Training data, evaluations and environments, built around real financial work.
Give the model material worth learning from.
Worked examples and training records built from real financial decisions, reviewed by practitioners.
Training and evaluation rights stay separate.
Illustrative specimen
Built from real deals, signed off by practitioners, delivered to the buyer’s format.
Follow an analysis as evidence arrives, a view changes and the work takes shape.
Reconstructed and practitioner-calibrated; not raw analyst-session replay.
Illustrative specimen
A practitioner-calibrated demonstration of how the work is done, not a record of a live session.
Examine the judgment behind an answer.
Assess financial judgment against fixed criteria, with evidence behind each score.
Criteria freeze before evaluation. Later outcomes never become the reward.
Illustrative specimen
The public principle is evidence-linked, decomposed, practitioner-reviewed assessment.
Carry the work through a changing case.
One agent carries a case across stages, with work and decisions that persist.
Information arrives as the case advances; each step uses only what was knowable then.
Illustrative specimen
117 graded tasks across 4 stages and 7 reasoning categories. Counts are the reference case as published; the per-stage distribution remains sealed.
A dated financial world, its tools and graders, ready to run inside your infrastructure.
Already have an environment? Receive the underlying data, tasks and reward components.
Illustrative specimen
Covenant headroom narrowed while revenue held flat. What broke, and when?
Training records, graders and evaluations arrive in your data formats. A runnable environment, where useful, runs inside your own infrastructure.
| Material | For | Delivery |
|---|---|---|
| Training data | Worked examples and training records | Versioned JSONL or Parquet, with schema documentation |
| Reference trajectories | Demonstrations, corrections and continuation pairs | Controlled reference artifacts with provenance |
| Sealed evaluations and graders | Evidence-linked, decomposed assessment | Scores with replayable transcripts and rubric versions |
| Long-horizon workflows | Staged financial work, from screening onward | Point-in-time datarooms and permitted tools |
| Packaged RL environments | Execution in the buyer’s evaluation or training perimeter | Runnable environments, or underlying components |
Run one sanitised case in your acceptance harness. Expand only if it adds measurable signal. No stack adoption required.
Talk to usComponent views are illustrative renderings. They reflect published methods and measured task totals where stated; the training-data and trajectory views are artifact sketches. For common questions, read the FAQ.