Skip to content
C.W.K.
Stream
Lesson 01 of 04 · published

Sixty-Seven of Seventy-Nine

~12 min · measurement, process, failure-analysis, bilingual

Level 0Wet Clay
0 XP0/36 lessons0/12 achievements
0/100 XP to next level100 XP to go0% complete

The Number That Started a Tool

A corpus of seventy-nine technical courses, each published in two languages. The English was written first, because it had to be: the code samples must match, the links must match, the syllabus must be one course rather than two. Then the same author wrote the Korean.

An audit read the Korean. Sixty-seven of the seventy-nine had the same defect. Not a typo class, not a formatting slip — the sentences kept their English words in English clause order and translated only the joints, so the Korean was grammar without vocabulary. Fluent enough to pass a spellchecker and wrong enough that a native reader stumbles on every third sentence.

Here is the part that turns a quality problem into an architecture problem. Three successive generations of explicit rules had been in force the whole time. "Do not translate — think in Korean and write." Every author could recite it. Every author agreed with it. Every author was trying to follow it. The rule got sharper each generation, and the corpus got worse.

What the Defect Actually Looks Like

You do not need Korean to see it, because the defect is structural rather than lexical. Look at what stayed English in a sentence that is supposed to be Korean: the nouns, the verbs, the adverbs — everything that carries meaning. What got translated was the connective tissue: the particles, the endings, the little grammatical hinges.

That inversion is the whole signature. A person writing Korean picks Korean words and arranges them in Korean order. A person translating picks the source sentence and swaps its skeleton, because the source is right there, already shaped, and re-shaping it costs more than dressing it.

Why This Is Not a Discipline Problem

The reflex response to a 67-of-79 failure rate is to conclude that people were careless, and to write a better rule. Test that theory against the population. A discipline problem has a specific shape: the failures cluster among people who did not know the rule, did not believe in it, or were under time pressure that made ignoring it rational. Here the failing set included the authors who could quote the rule verbatim, who had themselves argued for it, and who had no deadline at all.

When a failure survives knowledge, agreement, and motivation, you have measured the arrangement, not the people. And the arrangement was this: the author writing Korean had the finished English sitting in working memory, because they had just written it. They were asked to not use it. Asking someone to hold something in mind and not draw on it is asking them to forget on command — which is the one thing minds do not do on request.

A rule that fails among people who know it, agree with it, and are trying to follow it is not a rule problem. Rewriting it is the most natural response and the least useful one, because it treats knowledge as the missing ingredient when knowledge was never missing. The next question is not "how do we say this better" but "what were they holding when they broke it, and who handed it to them."

Code

The defect, and the same idea written as Korean·text
# BEFORE - published, and typical of the 67

  Mamba-family production 모델 하나 골라 (Jamba 1.5 Mini 가 친한 선택
  - open weights, well-documented). 같은 input 에서 Llama 3.1 8B 와
  quality 비교. specifically Mamba 가 어디서 이기는지 봐.

# Read the shape, not the words. Every technical noun and every
# discourse word stayed English: model, input, quality, specifically,
# open weights, well-documented. What got translated is mostly
# joints: 하나, 골라, 에서, 와, 어디서, 봐.
# The English discourse adverb 'specifically' sitting at the head of a
# Korean sentence is the tell - Korean does not front-load that way.

# AFTER - the same instruction, conceived in Korean

  Mamba 계열 중에 실제로 서비스에 쓰이는 모형 하나 골라. Jamba 1.5
  Mini 가 무난해. 가중치가 공개돼 있고 문서도 잘 돼 있으니까. 같은
  입력을 Llama 3.1 8B 에도 넣고 품질을 비교해봐. Mamba 가 어디서
  이기는지, 그 지점을 찾는 게 목적이야.

# Same facts, same links, same model names. Different skeleton.

External links

Exercise

Find a rule in your own process that people keep breaking. Write down three facts about the last five violations: did the person know the rule, did they agree with it, and were they under pressure. If the answers are yes, yes, and no, stop drafting a clearer version of the rule and instead write one sentence naming what the person was holding at the moment they broke it.
Hint
The useful question is about inputs, not intentions. Something was in front of them that made the forbidden move the cheapest available one — an open file, a pre-filled form, a default that was already correct-looking. Name that thing and you have named the fix.

Progress

Progress is local-only — sign in to sync across devices.
Spotted a bug or have feedback on this page?Report an Issue

Comments 0

🔔 Reply notifications (sign in)
Sign inPlease sign in to comment.

No comments yet — be the first.