The first RL benchmark for structured knowledge work
Train and evaluate LLM agents on real code review tasks — severity triage, queue prioritization, and actionable feedback generation — with deterministic grading and trajectory export for world model research.
Classify each PR's bug severity: critical, high, medium, low, or none.
Sort the review queue by urgency — security first, junior devs next.
Write actionable review comments targeting specific buggy lines.
Select a task above to start an interactive demo
States are structured text (code diffs, bug categories). Transitions depend on professional judgment, not physics.
Every episode exports (s, a, r, s') transitions in JSONL for training Knowledge-Work World Models.
No LLM-as-judge. Ordinal matching, Kendall Tau correlation, and 5-component weighted scorers.
Spam penalties, consistency checks, and decaying rewards prevent trivial gaming strategies.
| Benchmark | State Space | Transition | World Model? | Domain |
|---|---|---|---|---|
| MuJoCo | ℝⁿ (joints) | Physics sim | ✅ Dreamer | Robotics |
| Atari | Pixels | Game engine | ✅ MuZero | Games |
| TextWorld | Synthetic text | Game rules | ⚠️ Li et al. | Text games |
| SWE-bench | Code | N/A | ❌ Eval only | SE |
| CodeReviewEnv | Structured text | Professional judgment | ✅ KW-WM | Knowledge work |
POST
/reset
Start new episode
POST
/step
Take an action
GET
/state
Current state
GET
/health
Health check
GET
/export_trajectory
JSONL trajectory
GET
/docs
OpenAPI docs