Dead Archive → Topic Pipeline

Step 1 — Inventory (real script output, not a mockup)

$ python scripts/inventory_notes.py ./notes
扫到 4 个文件，入表 0 条（新增 0 条，跳过碎片 4 个）
  被跳过的（低于 --min-chars 200，不静默丢）：
    · 2025-03-ai-cs-rollout.md（99 字符）
    · 2025-05-pricing-notes.txt（75 字符）
    · 2025-08-archive-research.md（61 字符）
    · 2025-09-misc.md（8 字符）
  觉得丢多了就调小阈值：--min-chars 100
已写出: ./盘点表.csv

All four are short notes, so the default 200-character floor calls them fragments. Notice what the script does

not do: it does not silently drop them. It names every skipped file, and the floor is a flag:

$ python scripts/inventory_notes.py ./notes --min-chars 80
扫到 4 个文件，入表 1 条（新增 1 条，跳过碎片 3 个）
    · 2025-05-pricing-notes.txt（75 字符）
    · 2025-08-archive-research.md（61 字符）
    · 2025-09-misc.md（8 字符）

The CSV it produces (judgement columns left blank on purpose):

Step 2 — Tag by reusable argument, not by source

This is the step that decides whether the archive stays dead. Two notes, tagged two ways:

Now they share a tag. Source-tagging keeps them in separate drawers forever; argument-tagging puts them in the same drawer — which is the only place a cross can happen.

Step 3 — Cross

The matrix pairs tags and asks one question of each pair: what can I say now that neither note says alone?

Neither row exists in any single note. That is what crossing buys, and it is why the pipeline generates topics instead of waiting for inspiration.

Step 4 — Topic cards

Card 1 — 「AI 项目为什么失败：两个行业，一个原因」

Angle: the same failure mode shows up in customer service and in knowledge management — nobody fed it real business facts

• Evidence: 答错率 41% → 9%（补入 120 条业务问答后）; 收藏夹三年零产出（按来源归档，从未交叉）
• Backed by: S001, S004

Why it is yours: the four-hour detail of walking 120 Q&As one by one with the owner cannot be generated by anyone who did not sit through it

Card 2 — 「上线前不做一次真实数据对齐，等于什么都没做」

• Angle: the single shared predictor of failure across both projects
• Evidence: 41% → 9%; the archive that produced nothing until it was re-tagged by argument
• Backed by: S001, S002, S004

What this does not do

It generates topics, not finished articles. Crossing only helps if the archive has substance. And the inventory script reads plain text/markdown/csv — it is not a scraper for locked platforms.