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。
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。
Muir streams 10M tokens in a fixed 25 MB state, but stability, exact recall, multi-hop readout, and TPU throughput are four different claims.
用反事实任务对区分真实阻塞、透明拒绝和静默欠表现,把安全研究 Agent 的协作质量变成发布门。
Test whether safety framing changes agent performance with matched tasks, paired metrics, refusal transparency, and monitor-recall gates.
FlashProve 能把被检查 token 绑定到承诺的整数模型。采购和安全验收还要分别核对模型身份、服务链、质量安全与实现等价。
FlashProve can bind checked tokens to a committed integer model. This audit separates model identity, output integrity, safety, and implementation gap
Claude 完成了费马大定理的 Lean 形式化。本文区分命题范围、机器检查、专家审阅与独立复现,解释每层证据究竟能证明什么。
LifeVal 已能检查产物和状态变化,真实上线还要回答 Agent 是否有权行动、后果能否撤销,以及外部结果是否真正完成。
LifeVal tests personalized everyday tasks with artifacts and state diffs. Use a risk and verification ladder before an agent acts in the real world.