We build expert-grade environments for LLM reinforcement learning: real source materials, tasks that demand long-horizon tool use, multiple verifiable criteria, and complete reference trajectories from frontier models. Covering finance, law, medicine and general office work, delivered in Harbor format and ready to plug into RL training and evaluation.
We pick tasks from real business with a clear deliverable, professional judgement and a checkable standard. In three-model tests the investment-banking valuation averaged just 0.324 and a frontier model scored 0.14 blind on the quant engineering task — plenty of headroom for reward.
Original exchange PDFs, SEC registrants' 10-K/10-Q filings, real bid cost workbooks and a complete fictional litigation record. Models face the materials professionals face every day, not summaries prepared for them.
Every task requires searching, computing, writing code, editing files and cross-checking. Reference trajectories capture frontier models' full work — over 80 turns and millions of cumulative prompt tokens at the longest.
Every task has multiple independent criteria or a deterministic grader, with penalty items and capability gates. Hard-coded, empty and templated answers score nothing, and judge credentials are isolated from the agent container.
Each domain is split into subdomains, and each case has its own page. Click a domain to jump to its section of the RL environment overview.
Difficulty tiers come from multi-model test scores, not the author's own estimate. Criteria that fail to discriminate are replaced, and exploitable paths get penalty items.
Take original materials from real business and public disclosures, remove anything that gives away the answer, and keep the noise and conflicting definitions that real professional work contains.
Domain experts write the prompt along the real workflow: goals and constraints only, no method; tagged with capabilities, complexity and weakness labels.
Write multi-dimensional criteria or a deterministic grader, prove full credit is reachable with an oracle, and prove it cannot be gamed with empty, hard-coded and adversarial submissions.
Run several frontier models in real agent harnesses, set the difficulty tier from their mean score, and strengthen the criteria where they fail.
Every environment can be customised, with delivery in 2–10 days: full sample packages, batch authoring by domain or weakness label, multi-model baselines and RL platform integration support.