AIZEN

AIZEN is a reinforcement-learning environment for training agentic coding models on complex feature development. Each instance starts from a real repository at a pinned commit, masks the implementation of a target feature while preserving enough interface and behavioural context for the task to be executable, and grades the agent by running tests rather than by judging prose. Reward is deterministic and execution-derived: the requested feature must move from failing to passing while protected surrounding behaviour stays passing. Tasks are generated from repository scanning, dynamic tracing, masking, and post-verification — not hand-authored — and each bundle carries the provenance needed to reconstruct where it came from and which verifier contract grades it.

• Execution-verified reward: fail-to-pass and pass-to-pass test contracts computed from verifier output — no LLM judge, no reference transcript, no prompt-parroting credit.
• Reproducible containerized substrate: pinned repository commit, Docker environment metadata, and an explicit callable interface contract per task.
• Feature-level long-horizon work: L1 tasks preserve repository context for incremental feature work; L2 tasks require from-scratch implementation against the interface contract.