GoKawiilAI Art Training Rubric Failure Traced to Redundant Reward Signals
A team training a generative model to paint with code (using p5.brush) found their nine-signal reward rubric caused the model to plateau at flat, repetitive outputs. Investigation revealed four quality judges and prompt adherence were nearly identical measurements, while a code-length metric saturated early and stopped providing useful feedback, leaving only one signal actually driving learning. The fix replaced absolute 0-10 scoring with pairwise comparisons against a curated pool of 1,664 hand-rated reference images.