用任务结果、决策等价、约束保留和单个成功任务成本,验收 Agent 上下文压缩,而不是只看 Token 节省率。
Evaluate AI agent context compression with task outcomes, decision equivalence, constraint retention, and cost per successful task.
Claude-shaped science works when research problems, executable checks, and expert judgment fit the model's real capabilities.
用诱饵仓库、登录态差分、文件系统追踪、网络出口捕获和删除证据,在登录前审计 Coding Agent 的仓库数据流。
Audit coding agent data flows before login with canary repositories, auth-state diffs, filesystem traces, egress capture, and deletion evidence.
补丁半衰期把重写争论变成可复算决策:记录同类故障复发时间、修补成本、迁移成本,以及验证和回退能力。
Patch half-life turns rewrite debates into a measurable decision: track recurring failures, repair durability, migration cost, and rollback readiness.
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.
HANDBOOK.md 说明规则进入 Context 仍不足以形成控制。本文给出规则检索、执行前判定、提交门控与事后核验四层方案。