LAB / MONITOR WALL

实验室Lab

别读,看它跑。Don't read it. Watch it run.

运行记录Run log

FORM PREVIEW每一格屏幕是一次运行。内容就位后,这些占位会换成真实画面。Each screen is one run. Placeholders become real footage once content lands.

成功Passed部分Partial失败Failed

THE BENCH

十二个实验Twelve experiments

Q · R · E · P
  1. I洗掉一半Prune Half——
  2. II模型在看哪里Where the Model Looks——
  3. III偏好,还是长度Preference, or Length——
  4. IV第几代开始退化Which Generation Degrades——
  5. V化学数据里的错Errors in Chemical Data——
  6. VI重复的前缀Repeated Prefixes——
  7. VII循环的深度Depth Through Loops——
  8. VIII评分即对齐Scoring Is Alignment——
  9. IX想够了吗Thought Enough——
  10. X该不该自己决定Decide or Defer——
  11. XI离执行更近Closer to Execution——
  12. XII尚未命名Not Yet Named——

ARCHIVE

前期记录Earlier log

  1. 提示注入防御:从检测到最小权限Prompt-injection defence: from detection to least privilegeAI SECURITY NOTE
  2. Claude Code 项目能力隔离Per-project capability isolation for Claude CodeDESKTOP TOOL
  3. 本地 AI 交付控制面A local control plane for AI deliverySYSTEM PROTOTYPE