Model misalignment reporting needs replayable evidence, calibrated claims, independent access, and explicit closure criteria.
Engram 把条件记忆与条件计算分开。本文核验它移出 GPU 的内容、论文证据边界和部署评测方法。
DeepSeek Engram separates conditional memory from computation. See what moves to host memory, what evidence proves, and how to benchmark it.
DeepSeek V4.1 用 SWA 有界重放压缩持久 KV 缓存。本文拆解缓存命中时变化的状态、已有证据与上线前测试矩阵。
DeepSeek V4.1 cuts persistent KV storage with bounded replay. See what changes on a cache hit, what was tested, and how to validate it.
Web Bot Auth 只能验证签名身份。机器人访问治理还需要站点偏好、用途声明、名录分类、本地策略和可审计的边缘执行。
Web Bot Auth verifies a signing key. Robot access needs declared purpose, directory classification, local policy, and auditable edge enforcement.
KV cache 操作让预训练模型并发观察、思考和响应成为可能。机制已有实验，生产运行时仍在形成，状态治理将成为新的系统边界。
KV-cache manipulation lets pretrained LLMs observe, reason, and respond concurrently. The mechanism works; the production runtime is still emerging.
Astra 在不同 Harness 下得到 62.7% 和 99.9%。本文拆解分数语义、状态管理差异，以及速度和 token 数据的适用边界。
