{"version":"2026-09-27.1","editorialVersion":"2026-09-27.2","changes":[{"id":"C02","date":"2026-09-27","type":"资料维护","title":"DABstep区分已验证与未验证结果","before":"介绍页的早期结果容易被误作当前能力","after":"已验证榜单暂停新增验证；v2待发布","impact":"只引用具体结果状态和日期；没有取到可靠最新数值时保留原始入口。日期为本站核查日。","action":"使用数据分析证据时，先核对榜单验证状态；继续按本地业务门槛验收，不因榜单维护状态改变委派范围。","source":"E06","tasks":[{"id":"DATA-01","name":"核对订单收入，找出差异"},{"id":"DATA-02","name":"统一不同工具中的指标口径"},{"id":"DATA-03","name":"解释转化率变化，拆分人群组成影响"}]},{"id":"C05","date":"2026-09-27","type":"资料维护","title":"营销自测保留可复核性缺口","before":"官网宣称任务、评分与原始运行已公开","after":"本轮访问所链仓库返回404","impact":"自测结果附明显限制，不升级为独立验证；后续重查公开资产。日期为本站核查日。","action":"将营销供应商自测作为补充线索；验收标准不因自报分数而降低。等待可访问的任务和运行记录再评估支持力度。","source":"E14","tasks":[{"id":"GROW-01","name":"按人群制作有事实依据的内容"},{"id":"GROW-02","name":"让内容适配渠道和商业属性"},{"id":"GROW-03","name":"设计不同人群的激活实验"},{"id":"GROW-04","name":"核算增长增量和扣除成本后的收益"}]},{"id":"C01","date":"2026-09-22","type":"资料维护","title":"SWE-Bench Pro修订任务与运行协议","before":"旧公开集731题","after":"V2为642题，修正任务并锁定协议","impact":"开发任务保留基准版本。旧版与新版成绩分开解释，不自动认定能力回退或进步。","action":"在软件回执中记录V2与运行协议；与旧版比较前先检查任务集合和配置。","source":"E01","tasks":[{"id":"DEV-01","name":"修复重复提交导致的重复写入"},{"id":"DEV-02","name":"核对业务写入与事件通知是否一致"}]},{"id":"C03","date":"2026-06-26","type":"工作建议变化","title":"OSWorld 2.0补充长流程与跨应用证据","before":"主要参考单应用或较短操作任务","after":"新增108个长流程任务与整题/部分评分","impact":"企业交付页加入状态恢复、跨来源核对和最终验收要求，仍不宣称现场采用已验证。","action":"系统交接增加跨应用状态核对、部分失败与恢复检查；这是方法建议，不是现场成功率提升。","source":"E23","tasks":[{"id":"FDE-02","name":"连接业务系统，交付可恢复的流程"},{"id":"GROW-02","name":"让内容适配渠道和商业属性"}]},{"id":"C04","date":"2026-03-12","type":"工作建议变化","title":"失败诊断也有可引用的评测","before":"仅从是否完成观察Agent表现","after":"引入失败定位与根因分类证据","impact":"DEV-04增加失败轨迹审查；诊断增益与修复完成度分开展示。","action":"Agent失败后保留轨迹与根因判断，分别验收诊断是否正确、修复是否有效。","source":"E19","tasks":[{"id":"DEV-04","name":"评测工具Agent，定位失败与回归"}]}]}