Tianyu

A nano replica of Jev: parallel decisions, dynamic candidates, and an end-to-end training pipeline.

LLM-authored task-specific Jev harnesses with optional full-trajectory reward reflection and GEPA evolution.