Announce evaluator fixes and leaderboard update

[ahydchh]

No description provided.

[github-actions]

🤖 AI Code Review (gemini-3-flash-preview)

🇬🇧 English Analysis

1. Executive Summary

• Core Purpose: This PR is a documentation update to the project's main README files (English and Chinese). It announces critical updates regarding evaluator security, bug fixes in task validation, and a subsequent leaderboard refresh.
• Modified File Structure & Modifications:
  • README.md: Added a news entry dated 2026-09-16 detailing evaluator isolation improvements, validation fixes, and leaderboard updates.
  • README_zh-CN.md: Added the corresponding Chinese translation for the news entry.

2. AI Content Analysis

• Estimated AI Component: 5%
• Reasoning & Evidence: The content is highly specific to the project's internal logic (e.g., "candidate–evaluator isolation," "v1/v1-lite leaderboards," "Medal Score"). The phrasing follows standard professional changelog conventions. While an AI might have been used to polish the translation, the domain-specific nuance suggests human authorship or heavy human editing.

3. Engineering & Economic Assessment

• Engineering Reality Check: The PR describes a response to a production-grade engineering problem: benchmark integrity. Strengthening "candidate-evaluator isolation" is a non-trivial security requirement in LLM evaluation to prevent models from "cheating" or escaping the sandbox to manipulate scores. This addresses real-world edge cases where adversarial inputs could compromise leaderboard validity.
• Economic Value: High. For a benchmark project, the "product" is the data integrity. By fixing manipulated scores and improving isolation, the project maintains its reputation and utility. It reduces the "technical debt" of unreliable metrics.

4. Quality Assurance

• Verification & Testing:
  • frontier_eval Integration: No (This PR only contains documentation changes).
  • task_name: N/A
  • Execution & Dependencies: N/A. The PR does not introduce new code or environment requirements.
• Documentation Quality: High. The updates are concise, dated, and provided in both English and Chinese. The links to the leaderboard are correctly maintained. No spelling or grammatical errors were detected.
• Organizational Structure: Logical. The news items are appended to the top of the "News" section in chronological order.

5. Security & Privacy Check

• Sensitive Files: Clean.
• Absolute Paths: None detected.

---

🇨🇳 中文分析

1. 摘要

• 核心目的: 本 PR 是对项目主 README 文件（中英文版）的文档更新。主要发布了关于评测器安全性增强、任务校验漏洞修复以及随后的榜单数据更新的重要公告。
• 修改的文件结构与变更摘要:
  • README.md: 新增 2026-09-16 的新闻条目，详细说明了评测器隔离性优化、校验修复及榜单更新。
  • README_zh-CN.md: 新增对应的中文翻译条目。

2. AI 成分分析

• 预估 AI 含量: 5%
• 判断依据与证据: 内容高度依赖项目特定逻辑（如“候选代码与评分器隔离”、“v1/v1-lite 榜单”、“Medal Score”）。措辞符合标准的专业更新日志规范。虽然可能使用 AI 辅助润色翻译，但领域特定的细节表明其核心内容由人工撰写或经过深度人工编辑。

3. 工程与经济评估

• 工程现实检验: 该 PR 描述了对生产级工程问题的响应：评测基准的完整性。加强“候选代码与评分器隔离”是 LLM 评测中一个非平庸的安全需求，旨在防止模型通过“作弊”或逃逸沙箱来篡改分数。这解决了对抗性输入可能损害榜单有效性的实际边缘情况。
• 经济价值: 高。对于一个 Benchmark 项目，其“产品”就是数据的公信力。通过修复被篡改的分数并改进隔离机制，项目维护了其声誉和实用价值，减少了不可靠指标带来的“技术债务”。

4. 质量保证

• 验证与测试:
  • frontier_eval 集成: 否（本 PR 仅包含文档修改）。
  • task_name: N/A
  • 运行与依赖: N/A。本 PR 未引入新代码或环境依赖。
• 文档质量: 高。更新内容简洁明了，标注了日期，并提供了中英双语。指向榜单的链接正确。未发现拼写或语法错误。
• 组织结构: 符合逻辑。新闻条目按时间倒序添加到“News/新闻”栏目的顶部。

5. 安全与隐私检查

• 敏感文件: 未发现异常。
• 绝对路径: 未检测到。