Claw-Eval Model Card

这是可索引项目报告证据页：它保留 Claw-Eval Model Card 的源材料入口、机制线索和限制提醒；正文仍需 reader/editor 与 academic public-copy review 后才能当作最终结论引用。

Claw-Eval Model Card

1. Role in Self Evolve

用 human-verified tasks、rubrics 和 Pass^3 多次运行逻辑降低 lucky-run 偏差，适合作为可信 agent evaluation 样本。

2. Working Principle

300 tasks + 2,159 rubrics + Pass^3 + full-trajectory auditing。

3. Evidence Path

web GitHub page captured March 2026 evaluation logic and task split signals.

4. Teaching Use

Use this card to explain whether a repository is a runtime, benchmark, harness-evolution loop, memory/skill substrate, or resource index. The key reading path is: raw capture -> classification row -> public site card -> project report -> aggregate GitHub analysis.

5. Limits

未复现 leaderboard；许可证字段未在公开页面片段中确认。