Training a coding model to paint watercolours with TRL and OpenEnv

A coding model can be trained toward a visual style by judging what its programs draw. This experiment has the model write JavaScript for a watercolor-style rendering library; a browser renders the code, and scoring models evaluate the resulting image, including comparisons with a curated reference pool. The author reports that runs using the reference-based judge won more comparisons against that pool as training progressed. This supports learning the selected aesthetic preference, with the reference collection determining what the training rewards. Experiment.