What is a golden patch, and how do you construct one to correct an LLM-generated trajectory that failed during a distributed training run?
