Each sketch was rendered to PNG in a sandboxed Puppeteer environment, then updated with GRPO using a reward signal.
The initial nine-signal rubric plateaued at roughly 0.65 reward and produced similar clip-art-like flowers. Narreddi reports that several quality judges and prompt adherence scores were correlated at 0.85 to 0.95, while a code-length component stopped contributing useful gradient after about 30 steps.
The revised rubric used a 5% compile-and-brush-use gate, 5% length check, 30% HPSv3, and 60% pairwise judging against a hand-rated reference pool. The authors rated 1,664 images and selected 117 “love-tier” examples; the post also describes a 581-reference pool. It reports that the revised run reached the earlier plateau three times faster and reduced generated code from 13,500 to under 2,000 tokens.
The post does not establish independent evaluation, generalization beyond its watercolour task, or public availability of the trained model, code, or dataset.