The Traditional corpus needs a native reviewer for 14 contested terms — and two of them collapse different words into one string

[hotlong]

Filed by the PM seat from the corpus-wide survey in PR #192 (card #144). The ruling on #144 said the seat would decide what needs a native reviewer rather than let a generator card settle the house's Traditional vocabulary; this is that decision, with the measurement attached so a reviewer has something concrete to react to.

No fix proposed. Nothing here is dispatchable until someone who reads Taiwanese technical writing has looked at it. The unambiguous mechanical artifacts are split out to #194 and are not part of this.

Why this is not more of #144

#144 was decidable without a native reviewer because 許可權 failed a test that does not need taste: the corpus used it 435 times and used 權限 zero times, in a security-and-IAM corpus, which is prima facie a machine artifact rather than a style. The terms below do not fail that test. Every one is defensible as shipped. That is exactly what makes them a judgement rather than a bug.

The 14, by how much text they touch

Counts are of the source term in the zh-Hans corpus, which is the reliable number.

zh-Hans	n	files	twp ships	the alternative a reviewer would weigh
元数据	174	39	後設資料	中繼資料 / 詮釋資料
类型	67	20	型別	類型
发布	48	10	釋出	發布
审核	40	6	稽核	審核
优先级	20	9	優先順序	優先權 / 優先級
绑定	17	7	繫結	綁定
表达式	13	7	表示式	運算式 / 表達式
回调	11	2	回撥	回呼 / 回調
参数	8	5	引數	參數
映射	4	3	對映	映射
命名空间	3	2	名稱空間	命名空間
外键	3	1	外部索引鍵	外來鍵
局部	3	3	區域性	局部
抛出	3	1	丟擲	拋出
高级	3	3	高階	高級

元数据 alone is 174 occurrences across 39 files — on a site whose entire argument is about metadata-defined software, so it is the term a Taiwanese reader meets most often.

Two are worse than a vocabulary choice — they are collisions

These would still be worth a decision even if a reviewer liked every rendering above.

1. 视图 and 查看 both become 檢視. Verified directly:

视图  twp-> 檢視
查看  twp-> 檢視

72 occurrences of the noun and 16 of the verb land on one string, so 87 occurrences across 25 files carry two unrelated meanings — 「檢視企業客戶工單」 (verb, to view the tickets) sits beside 「一個高風險商機檢視」 (noun, a saved list/board view) in the same corpus. Worth noting for whoever decides: this corpus only ever means the application object (物件、欄位、檢視、流程), never a SQL VIEW, which is what would otherwise make 視圖 vs 檢視表 the hard part.

2. 运行时 becomes 執行時, which reads as the adverbial "at run time" — but the corpus overwhelmingly means the noun. 218 occurrences across 36 files, dominated by 可移植執行時 / 開源執行時 / 自託管執行時 / 基礎執行時 — a runtime you self-host, replace and audit. Four sentences use both senses at once and the result is visibly awkward:

• 「由執行時在執行時強制」 — enforced by the runtime at run time (ai-wrote-your-app-dare-to-merge)
• 「而不是可由客戶完整自託管的執行時。如果你的執行時必須存在於…」 (airtable-omni-vs-governed-ai-app-platform)
• 「開放自託管執行時，或者套件與自託管執行時混用」 (beyond-agentforce-copilot-open-runtime)
• 「而不是落進執行時。把它挪回執行時」 (eu-ai-act-runtime-audit)

A dictionary cannot split these by sense; 218 occurrences would have to take one term (執行時期 / 執行環境) or be edited by hand.

What a decision would look like

Three separable answers, and they do not have to be the same:

1. Leave the 14 as shipped, and treat twp as the house Traditional style. Cheapest, and defensible — the preset gets far more right than wrong, which is why OpenCC s2twp renders 权限 as 許可權 in every generated zh-Hant page (430 occurrences, 42 files) #144 kept it.
2. Override some subset, using the shared table PR gen-zh-hant: ship 權限/實例 and the straddled locative 裡 #192 builds. Each is one line plus a fixture case, so the cost is the decision, not the work.
3. Act on the two collisions only, leaving the vocabulary alone — they are the ones a native reader is most likely to notice, and the noun/verb merge in particular is not a matter of preference.

The seat's own reading, offered as input and not as a ruling: the two collisions are worth acting on and 元数据 is worth asking about because of its volume; the rest can wait indefinitely without harm.

Related

• OpenCC s2twp renders 权限 as 許可權 in every generated zh-Hant page (430 occurrences, 42 files) #144 / PR gen-zh-hant: ship 權限/實例 and the straddled locative 裡 #192 — the three terms already ruled and fixed, and the mechanism any of this would use.
• Five more s2twp over-substitutions are wrong in any register — 介面卡 is a network interface card, not an adapter #194 — the five over-substitutions that are wrong on their face and need no reviewer.
• Japanese renders "semantic layer" as 意味層, a literal calque rather than the term the category uses #159 — the same class of question in Japanese (意味層 → セマンティックレイヤー), which the seat was able to settle on observable usage. Recorded here as the contrast: that one had an evidence test, these do not.

Filed unassigned, finding. No files changed.

[hotlong]

Two additions from #194's implementation (c6ab4264), both of which belong on this card rather than that one.

1. 连接器 → 聯結器 — a 16th term the survey missed

Filed separately as #203 by the #194 dev, but it belongs in this table: 17 source occurrences, 34 in dist, one of them a post title. That would rank fifth here — above ten of the rows already listed. Contested rather than mechanical, so it was correctly left untouched.

2. 全局 → 全域 shipped under #194, and the seat's ruling there was too coarse

This one is mine to own. #194 was scoped to over-substitutions that are "wrong in any register", and I put 全局 → 全域性 → 全域 in that list. The 性 really is a mechanical artifact and it is gone. But the replacement I ruled is only right for two of the six occurrences.

The dev flagged it rather than shipping it silently, and the flag is correct. Two sites are the technical sense, where 全域 is plainly right:

錯誤卻是全域的 · 一個全域定時服務

The other four are the idiom for the whole picture, where 全局 is itself an ordinary Traditional word (綜觀全局), and 全域 reads as "the whole domain":

business-app-in-16k-tokens          …全域的。每…
ai-project-risk-assistant           …全域。
crm-ai-understands-customers        …全域。
beyond-agentforce-copilot-open-runtime  …全域。而這…

「專案經理就能看清全域」 is understandable and slightly wrong. It was 「看清全域性」 before, which was worse — so this is a net improvement, not a regression, and nothing needs reverting.

The dev offered three options and recommended A (ship as ruled), which was the right call for a card scoped to mechanical artifacts. But none of the three is the actual answer, and the mechanism to do better already exists in this repo:

D — pin the idiom by collocation, let the bare technical form convert. Identity entries on 看清全局, 掌握全局, 綜觀全局, 變成了全局; bare 全局 → 全域 otherwise.

That is exactly the shape of the LOCATIVE_LI whitelist #192 built for 里/裡 — repair behind a listed word, never a bare character — and it satisfies both senses without hand-editing a generated file. It was available and I did not think of it when I wrote the #194 ruling.

Recording it here rather than opening a card, because it is now a sense question and this is the register for those. Whoever answers this card's 14 (15 with 连接器) should answer this too; the fix is a handful of lines in the table src/lib/zhconvert.ts already has.

---

Generated by Claude Code