# Limitations

- This is a synthetic behavioral dataset.
- Repository contexts are structured abstractions rather than complete executable projects.
- Technical statements are designed to be plausible but have not all been verified by domain experts.
- Tool traces are intended actions, not observed executions.
- The generator uses controlled language patterns; semantic regularity may remain.
- Absence of exact or normalized duplicates does not prove complete semantic uniqueness.
- The test split holds out archetypes, but it is not a substitute for real-world evaluation.
- Models can still hallucinate, over-act, leak secrets, or make unsafe changes after training.