RL environments
Executable environments with a program as the reward signal, for training rather than only measuring.
Every task is already an environment: a working application, a proven exploit, a specification test, and a grader that runs all of it in an isolated sandbox. The reward is a process exit code, so there is no judge model to game and no preference label to disagree with.
A patch scores only when the exploit stops firing and the original specification still passes. That pairing is the whole design, and it is what stops a policy from learning to delete the endpoint: closing a hole by removing the feature is trivial, and it fails the second gate every time.
Delivered as environments you run, or behind a grading endpoint you call with a candidate patch and get a scalar back. Which one depends on whether you want the tasks inside your own training loop or the grading outside it.
What you get
- Environments packaged to run in your loop, or a grading endpoint you call
- Binary execution-verified reward, no LLM judge anywhere in the path
- Sandboxed execution with per-task isolation
- Fresh task generation on a cadence, so a policy cannot converge on a fixed set
Why this is not public
- Training against a public set teaches the set, and the leaderboard stops meaning anything
- A policy that sees the same tasks every epoch learns them rather than the skill
The method is fully inspectable either way. Browse every public task, exploit and reference patch before you talk to us.