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。
实验室 Agent 的能力上限由测量信号、通过条件与异常升级机制共同决定。MIT 六量子比特校准案例给出了一份可迁移的测量合同。
DseWiki 事件说明，只读网络应按外部副作用定义，而非按 HTTP 方法名称定义。本文分层审计公开档案、官方承认、站长证词与研究者推断。
KV-cache manipulation lets pretrained LLMs observe, reason, and respond concurrently. The mechanism works; the production runtime is still emerging.
Quantum calibration AI agents need every action tied to a measurable signal, a pass condition, and an escalation path. MIT's case shows the contract.
The DseWiki incident shows why read-only web access must be defined by side effects, not HTTP method names. Here is what the public evidence proves.
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。
Astra scored 62.7% or 99.9% on ARC-AGI-3 under different harnesses. Here is what changed, what the numbers mean, and what remains unproven.
Claude formalized Fermat's Last Theorem in Lean in 11 days. Here is what the 13-million-line proof and its verification stack establish.
持久 Agent 需要分别约束唤醒时机与行动权限。一次唤醒只允许观察，任何外部副作用都要重新校验授权。
