Which failure is missing?
Name the decision, evidence and failure pattern your current suite does not test. Describe the capabilities a successful agent would need.
FOR FRONTIER LABS
Your coverage gap.
A focused license.
Explore a finance dataset or environment license around your team’s missing capability.
For AI labs and developers: tell us what your existing data or suite misses. The proposed supply comes from finance software companies’ production workflows, corrections and outcomes. Start with one requirement and acceptance criteria, then inspect one approved dataset or environment package. Your acceptance, rejection or rework feedback guides what we source next.
Seeking initial buyers · Source-derived licensing is in development · No training gains are established
01 / DEFINE THE DELIVERABLE
An inspectable package.
A proposed license specifies the artifacts, versions and permitted uses. A dataset includes reviewed traces, corrections, context and supported outcomes. An environment adds runnable state, tools and grading; the components below describe that format.
- World & task
- Fictional source documents, initial state, task instructions and documented case provenance.
- Agent interface
- Tool schemas, observations, structured submissions, episode budgets and reset behavior.
- Grader & evidence
- Versioned scoring rules, reference checks, failure cases and the limits of each score.
- Release specification
- Reproducible configuration, family and split definitions, change history and agreed training or evaluation rights.
02 / BRING A CONCRETE REQUIREMENT
Three questions
before a pilot.
Your team can build tasks internally. A pilot should show useful additional coverage or preparation work it saves against that alternative.
How will your agent run?
Share the tool interface, environment contract, runtime constraints and intended use. Agree versioning, budgets, isolation and error reporting.
What would earn a pass?
Set acceptance criteria, acceptable alternative answers and independent case-family coverage. Agree the expert calibration needed before using policy scores as rewards. Review the exact package and explain which cases add coverage, need rework or miss the requirement; that feedback guides the next sourcing decision.
Today’s catalog is authored and synthetic: interactive tasks across ten finance segments, plus the connected accounting pipeline. Each task has an implemented deterministic grader; the segment exercises use bounded structured choices. Source-derived releases and independent expert calibration remain in development; training gains are not established.
Put one requirement on the table.
Send a non-confidential brief with your task, integration needs and acceptance criteria. We’ll discuss scope, permitted uses, pricing and delivery before an agreement.
Seeking initial evaluation partners. Starting a conversation creates no payment obligation.