Korean translationese (번역투 / calque) detection

src/features/translationese.js adds a deterministic, auditable signal for Korean calques — grammatical Korean that reads as translated-from-English. The stylometry + AI-lexicon signals catch structure and AI vocabulary; they do not catch lexical translationese (e.g. "커맨드 기둥" for "command pillars", "~에 의해" passives, "당신" for "you"). This catalog fills that gap.

Academic grounding

The concepts these signals operationalize — translationese, Toury's law of interference, Baker's translation universals, and Toral's post-editese, plus the Korean 번역투 research tradition (이근희, 김순영) — are documented with verifiable citations in research/ko-translationese-scholarship.md.

Design notes

• Precision first. Most of these constructions also appear in good native Korean (formal/technical prose especially). So the signal is density-gated: it only reports hot when both an absolute floor (≥4 hits) and a per-prose-sentence density (≥0.5) are met. A single "~에 의해" means nothing.
• Advisory, not a verdict. analyzeText surfaces it as translationese but deliberately keeps it out of the document hot decision, so it cannot regress benchmark false positives. SKILL, prompt, audit, and browser surfaces may show it as editing context only; they must not treat it as a score, gate, severity, benchmark, or authorship signal.
• ko-only for now (calques are language-specific).
• Each rule ships a before → after example (enforced by a unit test).

Catalog (before → after)

id	tell	before	after
noun-calque	직역 명사구 (pillar/layer 류)	세 가지 커맨드 기둥을 설치합니다.	핵심 커맨드 세 가지를 설치합니다.
dummy-subject	가주어 "그것은/이것은" (it is)	그것은 매우 중요하다.	매우 중요하다.
direct-address-you	"당신" 직접 호칭 (you)	당신은 이것을 설정할 수 있습니다.	이건 설정할 수 있다.
passive-e-uihae	"~에 의해" 피동 (by-passive)	작업은 에이전트에 의해 처리됩니다.	에이전트가 작업을 처리합니다.
have-overuse	"~을 가지고 있다" (have)	이 도구는 유연성을 가지고 있습니다.	이 도구는 유연합니다.
one-of	"~중 하나" (one of the)	가장 빠른 도구 중 하나입니다.	손꼽히게 빠릅니다.
provides	"~을 제공합니다" (provides)	다양한 기능을 제공합니다.	여러 기능을 쓸 수 있다.
as-follows	"다음과 같습니다" (as follows)	사용법은 다음과 같습니다.	사용법은 이렇다.
make-easy	"~하게 만들어 준다" (make it ~)	설치를 쉽게 만들어 줍니다.	설치가 쉬워진다.

Output shape

analyzeText(text, { lang: 'ko' }).translationese
// → { count, density, sentences, byRule:[{id,label,strong,count,example}],
//     hits:[...], hot, thresholds:{count,density} }

analyzeText(text, { lang: 'ko' }).koPostEditese
// → { schema:'koPostEditese.v1', lang:'ko', analyzed, skipReason,
//     paragraphCount, sentenceCount, eojeolCount,
//     metrics:{ lexical, endings, interference, rhythm },
//     paragraphs:[{ id, sentenceCount, eojeolCount, metrics }] }

koPostEditese.v1 is a Korean-only advisory payload for post-edited machine translation texture. It exposes raw counts and ratios so an editor can inspect literal pronouns, by-passives, light verbs, repetitive endings, suffix/rhythm patterns, and related Korean surface features. It is not calibrated evidence: callers must not treat these raw metrics as an authorship verdict, benchmark claim, score input, gate input, severity band, percentile, or baseline-derived finding. Audit surfaces may show the payload in a separate editing-hint section only; they must not mix it into deterministic severity rows.

Phase 4 calibration protocol / approval boundary

Phase 4 is a calibration and decision-report protocol only. It can measure whether translationese or koPostEditese.v1 correlates with Korean editing needs, but it cannot by itself authorize coupling. Any future coupling to score, hot, gates, severity, prompt/rewrite gates, benchmark pass/fail, or authorship language requires a separate product/spec decision and separately approved execution plan after the evidence package below is complete.

Required calibration package:

• Corpus strata: balanced Korean native-human controls, acceptable human translation/post-edit controls, raw machine-translationese samples, LLM Korean drafts, post-edited LLM drafts, and patina rewrite before/after pairs across blog, docs, marketing, academic/professional, forum/community, technical, and short-form UI/help genres.
• Labels: each item needs source/provenance, domain, length bucket, register, translation/editing status, and blinded human labels for needs_korean_edit, translationese_present, post_editese_present, and meaning_preservation_risk. Use at least two Korean-proficient reviewers with adjudication and agreement reporting.
• Metrics: report precision/recall/F1 for the labels above, false-positive rate on native-human and acceptable-human-translation controls, genre-stratified and short-text false-positive rates, confidence intervals, and representative wins/failures. If a separately approved offline prompt or rewrite experiment is proposed, its report format must also include human preference, MPS/fidelity regression, and edit churn.
• Ablations: measure translationese only, koPostEditese.v1 only, combined signals, raw counts versus normalized ratios, and lexical/endings/interference /rhythm groups independently.
• Decision thresholds: pre-register the minimum precision, maximum control false-positive rate, allowed MPS/fidelity regression bound, and minimum per-stratum sample sizes before looking at holdout results. Thresholds must be justified by holdout evidence, not by convenience on development examples.
• Rollback rules for any later approved coupling: coupled experiments must be feature-flagged or isolated; if false positives, MPS/fidelity regressions, browser/Node parity drift, or domain skew exceed the approved bound, disable the coupled behavior and keep advisory display.
• Deliverables: publish a corpus manifest schema, Korean editor labeling guide, offline experiment script/report format, representative wins/failures appendix, ADR template, and follow-on approval checklist. The ADR may complete inside Phase 4 only as advisory-only or reject-coupling. Prompt-context, rewrite-priority, score/gate, benchmark, or authorship-related options may be documented as follow-on proposals, but none may proceed without a separate product/spec decision and separately approved execution plan.
• No-coupling default: until a later approval explicitly names a coupling and cites completed corpus evidence, no translationese or koPostEditese.v1 metric may feed score, hot, gates, severity, z-score, baseline, percentile, prompt gates, rewrite gates, benchmark pass/fail, or authorship verdicts.

Limitations / next

• Calques are lexical; the structural stylometric classifier cannot learn them (proven separately). This is a rule catalog, like patina's pattern packs.
• The catalog is intentionally small and conservative; expand with corpus evidence and keep the density gate to protect precision.
• koPostEditese.v1 is not wired into hot, scores, gates, severity, z-score, baselines, percentiles, or benchmark decisions. Keep it advisory unless a separate post-Phase4 product/spec approval explicitly authorizes a narrowly scoped coupling.