A business workflow is more than a prompt. Each environment describes its own records, rules, documents, actions and correct outcomes—so an agent can be tested in realistic context.
Reusing the engine does not mean pretending every profession works alike. A new environment needs new knowledge and new evidence.
What exists, how records connect and which fields matter.
What a practitioner would check, and when the normal path changes.
What an agent may read, write and infer from permitted evidence.
What earns credit, what is unsafe and who has reviewed the judgement.
Finance operations is our live, empirically evaluated benchmark pack. Subsequent tracks expand into corporate operations, trade logistics, and wholesale RL environments for AI labs.
Linked suppliers, purchase orders, invoices, payments and bank entries. The agent must investigate mismatches, repair links, stage approvals and leave an immutable, tamper-evident trail.
Purchase requisition approvals, supplier onboarding, and tiered managerial sign-offs for corporate spend management.
Movement of goods, international bills of lading, customs tariffs, and multi-currency cross-border reconciliation.
Native adapter for Prime Intellect's `verifiers` hub, enabling RL training of reasoning models on enterprise tasks.
Gym-compatible environment with fast in-memory SQLite resets for PPO/DPO training at thousands of steps/sec.
Every pack needs clean data, correct planted cases, a solvable task, answer-safe tools, shortcut checks and practitioners who can challenge its realism. A finance result does not automatically transfer to another domain.
Read the benchmark method