Skip to content

v0.2.0 — 配对与证据完整性升级

Choose a tag to compare

@coolwkx coolwkx released this 19 Aug 06:51
· 7 commits to main since this release
475dbe6

Agent Eval Lab v0.2.0

  • 使用 taskId + repeatId + seed 严格配对,拒绝重复和孤立样本
  • 成功率与 McNemar 使用同一批有效配对
  • verification 证据必须与成功 tool_result 的产出一致
  • 自动测试 10/10、合成演示通过、依赖漏洞 0

附件是明确标注的合成示例报告,只用于复现评测流程。