Balancing Label Quantity and Quality for Scalable Elicitation — LessWrong