Open-source recipe trains coding model to paint watercolours

Sergio Paniego reproduced Surya Narreddi's watercolour-painting language model with TRL and OpenEnv, publishing all artifacts. The model writes JavaScript via p5.brush, limited to 10 methods. Rewards combine HPSv3 and Qwen3-VL-30B-A3B-Instruct pairwise judging against a hand-rated pool. Three reward mixes were trained and compared.