Running pose_fit.py on the new master image crashed with a NaN-to-int error; investigation revealed SuperAnimal keypoints were often wrong or outside the animal's silhouette on illustrated (non-photo) input.
The human: Developer looked at the bent-pose render and said the hind legs looked like garbage, not proper legs.
Built pipeline from pet photo to procedural walking rig with zero foot-slip →Real photo generations came back wrong: a mask compiled onto half the head made a shepherd look like a rottweiler, repeated runs on the same photo gave inconsistent genomes (one gave a dog a cat-like tail) because temperature was 1, and even after fixes the result still wasn't 'pretty' — the parametric compiler could reproduce Honey's proportions almost exactly but couldn't render fur, so a Maine Coon came out looking like a bare skeleton.
Turned a hardcoded tamagotchi into a AI pet generator →