VHD-Play builds agentic reinforcement-learning environments in the opposite order from typical pipelines. It first samples and solves a mathematical model, then uses a frozen corpus-grounded setter to turn the solved decision process into stateful tools. The same solved model supplies the environment dynamics and the trajectory-scoring reference, so the outcome signal is defined before the environment exists, not patched in afterward. The pipeline produced 3,300 environments at a few cents each.
Training Qwen3.6-35B-A3B on three mechanism families raised its mean agentic score from 0.204 to 0.815 on a five-family diagnostic. Gains transferred to held-out instances, eight unseen mechanism families, and external benchmarks for general function calling, travel planning, and 365-day e-commerce. On E-Commerce Bench the trained model completed every run without bankruptcy and exceeded Qwen3.7-Max. A comparison of written-out versus stateful versions showed most of the learnable gap came from stateful interaction, not underlying problem solving.