Every term in i18n/ko/glossary.json is marked "provisional pending native review", and #3280 mentions that only Portuguese had a reviewer lined up. I'm a native Korean speaker and compared the generated Korean pages with the English source. Below are only the problems that come from the inputs (glossary.json / instructions.md), not one-off sentences. Paths are under i18n/ko/pages/, and line numbers refer to those Korean files on main at 2af9e30.

AI disclosure: I used Claude Code to compare 45 of the 52 Korean pages with the English source and collect examples. I read and confirmed every example below myself, and the proposed renderings are my choices. advanced/ was not reviewed.

1. capability → 기능 collides with "feature"

The glossary note already anticipated this collision; the pages confirm it.

• handlers/dependencies.md:148 "선언되지 않은 기능은 …" (capability) and :150 "독립적인 샘플링 및 루트 기능은" (features): two lines apart, and the English contrast is lost.
• client/callbacks.md:73 ## A callback is a capability → ## 콜백이 곧 기능, which reads as "a callback is a feature".
• whats-new.md:192: the English italicizes capabilities to separate them from features, but the next paragraph uses 기능 for features too.

Proposal: render capability (what a side declares it supports) as 지원 기능, and keep 기능 for feature. Example: ## 콜백 등록이 곧 지원 기능 선언.

2. era has no entry, and "2025년대" is wrong

• handlers/sampling-and-roots.md:41 "On 2025-era connections" → "2025년대 연결에서". -년대 means a decade (as in 2020년대), so "2025년대" / "2026년대" is not a valid expression. It occurs 4 times (also servers/media.md:123, client/session-groups.md:72).
• Elsewhere the same word is 시대 (protocol-versions.md), 세대 (handlers/multi-round-trip.md) or 시절 (deprecated.md).

An era is a group of protocol revisions, not a period of time, so any time word misleads in Korean. Proposal: no noun for era; name the versions instead.

• 2025-era → 2025-11-25 이하 버전
• 2026-era → 2026-07-28
• pre-2026 → 2026-07-28 이전 버전
• a generic "protocol era" → 프로토콜 버전
• never -년대, -년, 시대, 시절 or 세대

3. Revision strings read as calendar dates

• "deprecated at 2026-07-28" became "2026-07-28에서 지원 중단 예정" (handlers/sampling-and-roots.md:49, handlers/dependencies.md:150). This reads as "support ends on July 28, 2026", while the English means "deprecated as of the 2026-07-28 revision".
• The term is also shortened to 지원 중단 ("호출하면 지원 중단 경고가 발생합니다", handlers/sampling-and-roots.md:43), which reads as "already unsupported".

Proposal:

• In instructions.md §4, when a revision string is the point in time, write 사양부터: "2026-07-28 사양부터 지원 중단 예정입니다".
• In the deprecated note, say that "deprecation warning" is 지원 중단 예정 경고 and that the term is never shortened.

4. elicitation → 엘리시테이션 is opaque

The glossary calls this the least settled term. The transliteration carries no meaning for Korean readers, and the candidate 유도 reads as "nudging the user".

Proposal: 사용자 입력 요청, written 사용자 입력 요청(elicitation) at first mention.

• form elicitation → 폼 입력 요청
• URL-mode elicitation → URL 모드 입력 요청
• push elicitation → 푸시 입력 요청
• code such as ctx.elicit() and elicitation/create stays as is

5. Terms with no entry, rendered differently across pages

Counts are approximate (grep over i18n/ko/pages).

The argument, parameter, annotation and docstring proposals follow the Python documentation's own Korean glossary (https://docs.python.org/ko/3/glossary.html).

"wire" (와이어 ×35, 전송 구간, 통신) needs a rule by context rather than one word:

• on / across the wire → 실제로 주고받는 메시지에서
• wire name, key or format → 프로토콜상 이름 / 키 / 형식
• the wire as the channel (stdout under stdio) → 통신 채널
• the verb "wire" → 연결하다 (already right)
• never 와이어

6. Gaps in instructions.md

• Requirement strength. There is no mapping, and "should not" came out at must-not strength ("-해서는 안 됩니다", "-면 안 되는지", "-지 않아야 합니다") in deprecated.md:155, client/identity-assertion.md:119 and handlers/sampling-and-roots.md:13. Proposal: must → -해야 합니다, should → -하는 것이 좋습니다, should not → -하지 않는 것이 좋습니다, may → -할 수 있습니다.
• Headings. §1 bans -나요? / -습니까? headings, but other sentence endings appear: "루트: 어디에 두어야 할까" (handlers/sampling-and-roots.md:28), "## 무엇을 발생시킬 것인가" (servers/handling-errors.md:78), "## 원할 때까지는 비용이 들지 않습니다" (run/opentelemetry.md:38). Proposal: say that a heading never ends in a verb ending.
• -ㅂ시다. It appears 3 times in servers/handling-errors.md (e.g. "바꿔 봅시다."). §1 could list it next to -십시오.

If this is useful and you assign me, I can open a PR with the glossary.json / instructions.md changes. I also have a list of one-off mistranslations (e.g. "crash" → 충돌 in client/index.md:142) that I can file separately.