Harder claims tasks
Longer rulebooks to search, deadlines that count days differently by state, facts that change mid-claim, legal notices and multi-claim queues.
Train · AI labs and RL teams
RL environments for AI agents, with verifiable rewards.
RL environments with verifiable rewards score an agent by checking the system's final state against written rules, not by asking a model to judge a transcript. AutoStack environments reset from a seed in about 6 ms, and do-nothing and shortcut agents fail every task. Labs commission private environments to a written brief.
Rewards from state checks
15 tasks, 3 seeds, 45 rewards.
Each episode resets from its seed, the agent acts, and the grader turns the final state into a reward. A shortcut can't score: graders are checked against agents that do nothing or game the task.
episodes graded45/ 45
Business workflows
Why train on business systems instead of puzzles?
Agents deployed at work make decisions about money, authority and exceptions. Our claims environment trains exactly that, against a written manual, with consequences the grader can measure.
- Deductible and limit math
- Reserves set before payment
- Payment authority and escalation
- Duplicate invoices
- Fraud (SIU) holds and act-or-escalate twins
- Instructions injected into a claimant's document
grade(episode) → final state ✓ → audit log ✓ → reward 1.0
do_nothing_agent → 0 / 45 shortcut_agent → 0 / 45
What you get
What makes an environment trainable
- resetDeterministic resetsEvery episode is built from a seed in about 6 ms (p95 8 ms) and starts from the same state.
- rewardRewards from state checksThe grader reads the final state and the audit log, and scores them against written rules.
- robustGraders checked against shortcutsDo-nothing and shortcut agents fail every task. Rejected attempts still count.
- toolsA real tool interface15 tools (6 read, 9 write) served over MCP and backed by a real database.
- variationSeeded variationEach task runs on 3 seeds, so a policy has to generalise across details, not memorise one case.
Custom environments
Can you build an environment for our workflow?
Built to your brief · priced by quote
A private environment, task set and graders for your workflow.
Written against your operating rules and delivered as a container, with rewards from state checks.
Send an environment brief- environment · container
- tasks · written to your brief
- graders · state checks
- seeds · reproducible resets
- tools · served over MCP
In development
Coming next for training
Private held-out task sets
Sealed tasks we run ourselves and never publish, so scores stay meaningful.
Run in your own tools
Export tasks to Harbor, Inspect and OpenEnv, including for RL training.
Full roadmap → Roadmap items are plans, not commitments.
FAQ
Questions from labs.
What makes a reward verifiable?
The grader computes it from the final state of the system and the audit log, against written rules. Rerunning the same task on the same seed produces the same state and the same score.
Can the agent game the grader?
Graders are checked against do-nothing and shortcut agents; both fail every task. Attempts the system rejected still count against the agent.
How is an environment delivered?
A private environment, task set and graders built to your written brief, delivered as a container. Priced by quote.
Do you only build insurance environments?
Insurance claims is the environment live today. Others are built to a brief with a design partner; tell us the workflow.