translate

name	translate
description	For translating between Korean and English in Lemma, QA-ing existing translations, or triaging which prose needs rework. Use when adding the counterpart locale to new content, when reviewing for awkward calque, or when the user flags translation quality.

Translate

Lemma is bilingual at the term level. Korean is not a translation — it is a native voice carrying the same insight. This skill is for getting the counterpart locale past the same content bar.

Register

• Technical but readable, with mild literary flair where the prose calls for it. The hook is allowed to be vivid.
• Use ~다 / ~이다 declarative throughout prose. No ~합니다 (the textbook-formal tone is wrong here — it signals school, not curiosity).
• Avoid academic stiffness. Avoid corporate Korean.
• The English voice is direct, slightly literary, occasionally aphoristic ("Logs aren't just inverse exponents. They are the bullshit detector for any story told in compounding."). The Korean must match that energy with Korean rhythms — not imitate the clause structure.

Pitfalls (real examples from this repo)

Direct calque from English produces awkward Korean. Some patterns to watch:

Calque (avoid)	Native (prefer)
이 모듈은 X 가 소비한다	이 모듈은 X 에서 쓰인다
당시 약 $41 을 보냈다	당시 약 $41 어치 — 를 보냈다
결과가 주어지면 단계 수를 센다	결과를 주면 몇 번 곱했는지를 센다
미쳐가던 천문학자들의 시대	속이 타들어 가던 천문학자들의 시대였다
더하기 같음 / 곱하기 같음	덧셈으로 같은 양 / 곱셈으로 같은 비율

Math notation stays in source form

• $\log_{10}$, not 로그₁₀
• e, not 자연상수 (the term e itself is taught; don't translate it)
• \Sigma, \int, function names — Latin / symbolic
• The exception: when the symbol is the Korean term being defined, define it via the glossary, not by translating in prose

All math goes inside $...$ (inline) or $$...$$ (block) using KaTeX-renderable LaTeX. A Korean label inside math wraps in \text{}: $g_{\text{달}} \approx 1.62$. Currency in prose escapes: \$100, never bare $100. Full rule in CONTRIBUTING.md.

Term policy

When an English term has a settled Korean form, use it: 복리 (compound), 로그 (log), 지수 (exponent), 거듭제곱근 (n-th root), 밑 (base).

When it doesn't, keep the English: float32, softmax, SPY, PyTorch. The <Term> hover shows the counterpart anyway, and forcing a Korean neologism (부동소수점32?) reads worse than the English everyone uses.

Code stays English: variable names, function names, comments. Always.

Term consistency (mechanical)

Every <Term id="x"> in prose requires:

• src/data/glossary/x/en.md — frontmatter term: + body
• src/data/glossary/x/ko.md — frontmatter term: + body
• The id appears in src/data/glossary/_order.json (else it won't render in the glossary section)

Verify with:

npm run validate:glossary --prefix apps/web

Triage workflow

When the user says "review Korean translations" or similar:

1. Survey all Korean prose surfaces:
   • apps/web/src/views/*.tsx (page prose)
   • apps/web/src/data/glossary/*/ko.md (glossary entries)
   • apps/web/src/data/{applications,modules}.ts (card titles + hooks)
2. Score each passage on calque ↔ native (1–5). Don't grade individual words; grade the paragraph rhythm.
3. Prioritize the 1–3 worst, with one-line diagnoses ("English clause order leaks through", "academic register", "term inconsistency with x/ko.md").
4. Use Opus for the rewrite pass on tough passages. Sonnet/Haiku tend to produce smoother but blander Korean and over-formalize.
5. Verify the rewrite preserves meaning, not clause structure. The English and Korean versions are allowed to diverge in how they say it as long as they converge on what the reader walks away knowing.

QA per file

For a single ko.md or ko.* surface, check:

• Does any sentence read like English-shaped Korean? Rewrite.
• Any ~합니다 mid-prose? Convert to ~다.
• Math notation in Latin form?
• Are technical terms consistent with the glossary entries?
• Does the rhythm feel native — short Korean sentences where the English was short, long where it was long? Not slavishly — just rhythmically.

When translating new content (en → ko or ko → en)

1. Read the entire page first. Translation is content-shaped, not sentence-shaped.
2. Translate the hook last. It's the most rhythmic prose on the page; it benefits from doing the rest first.
3. For the arc sections, preserve the function of each clause (this clause sets up, this clause delivers, this clause pivots). Don't preserve the order if the target language wants it differently.
4. For exercises and solutions, preserve the math literally. The prose wrapping the math can flex.
5. Check term ids — every <Term id="x"> in the source must resolve in the target's glossary. Add missing locales as part of the same pass.