基准污染需要评测侧复现门,而不只是一份训练数据声明

用四套分数、反事实任务和运行轨迹,把模型榜单从单一数字改造成可审计的采购证据。

Administrator Administrator Published on 2026-11-11

Benchmark Contamination Needs an Evaluation-Side Reproduction Gate

A practical gate for separating reproducible capability, grader loopholes, variant robustness, and unresolved contamination in AI leaderboards.

Administrator Administrator Published on 2026-11-11

California DROP 复查指南:把经纪商删除状态变成证据链

首轮处理窗口已过。按经纪商核对匹配、删除、豁免与待处理状态,并记录 2027 年处理周期变化和证明边界。

Administrator Administrator Published on 2026-11-11

托管 Agent API 可移植性:路由同名之后,审计控制面

迁移长程 Agent 前,逐项核对状态、事件恢复、权限、凭据、沙箱与数据边界,避免把资源同名误判为兼容。

Administrator Administrator Published on 2026-11-11

Managed-Agent API Portability: Audit the Control Plane, Not the Route Names

Audit state, event recovery, permissions, credentials, sandboxes, and data boundaries before moving a long-running agent between managed APIs.

Administrator Administrator Published on 2026-11-10

Meta Muse Code 运行时:事件日志、崩溃恢复与 Exactly-once 缺口

Muse Code 已退出 beta 并更换主力模型,但公开运行时证据仍未关闭 exactly-once 缺口。本文用故障注入审计其事件日志与恢复合同。

Administrator Administrator Published on 2026-11-09

Claude Code 权限文件也是敏感数据:审计从命令批准到 Git 提交的数据流

一份面向开发与安全团队的现场手册:检查 Claude Code 权限文件、Git 历史和构建制品,并按正确顺序处置泄露凭据。

Administrator Administrator Published on 2026-11-09

Claude Code Permission Files Are Secret-Bearing Artifacts

Audit Claude Code permission files as sensitive data across command approval, local settings, Git history, build artifacts, and credential rotation.

Administrator Administrator Published on 2026-11-09

手机 Agent 的进程内 Runtime:PhoneBuddySDK 双时间点审计

以两个代码快照审计 PhoneBuddySDK:进程内 Runtime 解决移动边界,代码修复与商店证据必须分开验收。

Administrator Administrator Published on 2026-11-08

Muir 线性注意力:1000 万 Token 证明了什么,又没有证明什么

拆开 Muir 的长流稳定、精确召回、多跳读出与 TPU 吞吐口径,避免把四组实验合并成通用 1000 万 Token 上下文结论。

Administrator Administrator Published on 2026-11-08
Previous Next