16. Refine the labels, not just the model
Improve labels through review queues, adjudication, guidelines, active learning, and disagreement analysis. Better labels often beat bigger models, especially when the task is subtle or human judgment is inconsistent.