According to OpenAI, the organization has developed detailed rubrics in collaboration with original paper authors to evaluate replication attempts of 20 papers. These rubrics contain 8,316 specific requirements assessed by an LLM judge, ensuring precise evaluation of replication fidelity.