A Reader Noticed Before We Did
One day a reader in Taiwan sent a short message: 「There are a few Simplified characters in this article.」 We checked, and it was true. Right in the middle of the Traditional Chinese body sat words like 「软件」 and 「设置」.
Going back through that month's records, we found six articles with the same problem. Five of them had gone unnoticed by anyone on the team until a reader pointed them out.
The Cause Was Surprisingly Simple
Our bot writes Traditional and Simplified versions in one pass, so habits from the Simplified sentences leaked into the Traditional side. It happened most often with words that look alike, such as:
• 軟體 / 软件
• 設定 / 设置
• 資料 / 数据
The review bot running in our Windows environment only checked whether the meaning was right. It never looked at the shape of the characters.
The Rule Was Just Three Lines
1. Scan the Traditional (zh) body for Simplified-only characters.
2. If even one appears, stop publishing and report only the line number.
3. Do the reverse for the Simplified (cn) body: look for Traditional-only characters.
The key was that we did not say 「fix it for me」 but 「find it and tell me only the line number」. When the review bot fixes things itself, it ends up touching other characters. Fixing goes back to the writing bot, and the review bot only searches.
Results
| Item | Before | After |
|---|---|---|
| Mixed-character articles per month | 6 | 0 |
| How it was found | Reader report | Blocked automatically before publishing |
| Added time | None | About 10 seconds per article |
What We Learned
In multilingual work, the dangerous mistake is often not a wrong translation but one inconspicuous character. Checking meaning and checking glyphs are different jobs, so they should be assigned separately. We copied the same three lines as they were into the Mac mini's overnight batch, and no reader reports have come in since.