仅凭分数不能证明发现：审计 AI 研究代理的发现认证协议
原标题：Scores Alone Do Not Prove Discovery: The Discovery Certification Protocol for Auditing AI Research Agents
AI 摘要
论文提出 Discovery Certification Protocol（DCP），用于审计 AI 研究代理的发现声明，将结果声明转化为可执行的恢复测试与反馈测试。DCP 包含 Gate 1 密封评估、Gate 2 匹配代理恢复测试和可选 Gate 3 反馈效应测量，核心要求零恢复见证和有限样本恢复概率上界。在 SQLite 优化和虚拟催化剂控制两个受控审计中，96 个 episode 均零恢复，配对研究得到 30 次真实恢复和零中性恢复，并通过 60 对零假设研究。确定性、无 LLM 的验证器可从冻结证据复现决策。
正文节选
Scores Alone Do Not Prove Discovery The Discovery Certification Protocol for Auditing AI Research Agents Abstract AI research agents combine prior knowledge, public sources, and experimental feedback to produce useful results. The Discovery Certification Protocol (DCP) turns claims about these results into executable recovery and feedback tests. Gate 1 validates useful improvement on sealed evaluation. Gate 2 gives matched agents the registered starting information and observed Web content while