patina-cli 3.11.0 → 4.0.0

package/artifacts/rebaseline-2025/intake.example.jsonl

@@ -0,0 +1,2 @@
+{"sample_id":"ko-public-blog-ai-001","language":"ko","class":"ai-like","register":"blog","model_family":"gpt-family","provider":"fixture","model":"fixture-gpt-family-2026-05","generated_at":"2026-05-21","prompt_id":"rb25-ko-blog-001","decoding":{"temperature":0.7,"top_p":0.9},"postprocess":{"editing_pass":"none"},"redistribution":"repo-ok","text":"이 가이드는 팀이 더 명확한 메시지를 만들 수 있도록 단계별 방법을 제공합니다."}
+{"sample_id":"ko-public-productdoc-human-001","language":"ko","class":"natural-human","register":"product-doc","model_family":"human-reference","provider":"fixture","model":"maintainer-written-reference","generated_at":"2026-05-21","prompt_id":"rb25-ko-product-doc-001","decoding":"not-applicable","postprocess":{"editing_pass":"maintainer copyedit"},"redistribution":"repo-ok","text":"설정값은 프로젝트 파일에서 먼저 읽는다. 값이 없을 때만 기본값을 쓴다."}

package/artifacts/rebaseline-2025/intake.local.example.jsonl

@@ -0,0 +1,25 @@
+{"sample_id":"ko-human-control-01","language":"ko","generated_at":"2026-05-21","decoding":"not-applicable","postprocess":{"editing_pass":"none"},"redistribution":"hash-only","text_hash":"sha256:eed8a722a97e18c4f151faa694c8126e85bab85ec54d9d9f8b924fc7bb4f9c7a","source_review":{"status":"template-placeholder","rationale":"Replace text/hash locally before using as evidence; do not commit raw text unless redistribution is cleared."},"class":"natural-human","register":"academic-summary","model_family":"human-reference","provider":"local-human-control","model":"maintainer-or-approved-human-source","prompt_id":"rb25-ko-academic-da-001","reviewer_notes":"Native Korean control candidate; keep raw text local unless explicit fixture permission exists."}
+{"sample_id":"ko-human-control-02","language":"ko","generated_at":"2026-05-21","decoding":"not-applicable","postprocess":{"editing_pass":"none"},"redistribution":"hash-only","text_hash":"sha256:58902326131c45cd5a44f65da5e8eaa71e150f350d18827ada9bc82fd7dc8a0c","source_review":{"status":"template-placeholder","rationale":"Replace text/hash locally before using as evidence; do not commit raw text unless redistribution is cleared."},"class":"natural-human","register":"academic-summary","model_family":"human-reference","provider":"local-human-control","model":"maintainer-or-approved-human-source","prompt_id":"rb25-ko-academic-da-001","reviewer_notes":"Native Korean control candidate; keep raw text local unless explicit fixture permission exists."}
+{"sample_id":"ko-human-control-03","language":"ko","generated_at":"2026-05-21","decoding":"not-applicable","postprocess":{"editing_pass":"none"},"redistribution":"hash-only","text_hash":"sha256:3b970b4c6f789ce5e18d5adb74c0175a1cc595ee69bc7cc4426cbf7fd14b0bc1","source_review":{"status":"template-placeholder","rationale":"Replace text/hash locally before using as evidence; do not commit raw text unless redistribution is cleared."},"class":"natural-human","register":"product-doc","model_family":"human-reference","provider":"local-human-control","model":"maintainer-or-approved-human-source","prompt_id":"rb25-ko-product-doc-001","reviewer_notes":"Native Korean control candidate; keep raw text local unless explicit fixture permission exists."}
+{"sample_id":"ko-human-control-04","language":"ko","generated_at":"2026-05-21","decoding":"not-applicable","postprocess":{"editing_pass":"none"},"redistribution":"hash-only","text_hash":"sha256:4bad105fb55795f507737ac0ab5092476442cb0f5a98993d8a0709e2103da73d","source_review":{"status":"template-placeholder","rationale":"Replace text/hash locally before using as evidence; do not commit raw text unless redistribution is cleared."},"class":"natural-human","register":"product-doc","model_family":"human-reference","provider":"local-human-control","model":"maintainer-or-approved-human-source","prompt_id":"rb25-ko-product-doc-001","reviewer_notes":"Native Korean control candidate; keep raw text local unless explicit fixture permission exists."}
+{"sample_id":"ko-human-control-05","language":"ko","generated_at":"2026-05-21","decoding":"not-applicable","postprocess":{"editing_pass":"none"},"redistribution":"hash-only","text_hash":"sha256:f13ef06a1e276cd9257f06411206eb3c55db0fa96cc0d373c0d234d7277d5375","source_review":{"status":"template-placeholder","rationale":"Replace text/hash locally before using as evidence; do not commit raw text unless redistribution is cleared."},"class":"natural-human","register":"blog","model_family":"human-reference","provider":"local-human-control","model":"maintainer-or-approved-human-source","prompt_id":"rb25-ko-blog-001","reviewer_notes":"Native Korean control candidate; keep raw text local unless explicit fixture permission exists."}
+{"sample_id":"ko-human-control-06","language":"ko","generated_at":"2026-05-21","decoding":"not-applicable","postprocess":{"editing_pass":"none"},"redistribution":"hash-only","text_hash":"sha256:b171223e21ee621901f4b3478e70ddf8d8624fdc71376f200605c4b08b5a8556","source_review":{"status":"template-placeholder","rationale":"Replace text/hash locally before using as evidence; do not commit raw text unless redistribution is cleared."},"class":"natural-human","register":"blog","model_family":"human-reference","provider":"local-human-control","model":"maintainer-or-approved-human-source","prompt_id":"rb25-ko-blog-001","reviewer_notes":"Native Korean control candidate; keep raw text local unless explicit fixture permission exists."}
+{"sample_id":"ko-human-control-07","language":"ko","generated_at":"2026-05-21","decoding":"not-applicable","postprocess":{"editing_pass":"none"},"redistribution":"hash-only","text_hash":"sha256:cee3cb6aa869a48288b2c513fcc9cc8c9bbf5b6cc3b4e2c4a78cf2e495f3d797","source_review":{"status":"template-placeholder","rationale":"Replace text/hash locally before using as evidence; do not commit raw text unless redistribution is cleared."},"class":"natural-human","register":"chat-update","model_family":"human-reference","provider":"local-human-control","model":"maintainer-or-approved-human-source","prompt_id":"rb25-ko-chat-update-001","reviewer_notes":"Native Korean control candidate; keep raw text local unless explicit fixture permission exists."}
+{"sample_id":"ko-human-control-08","language":"ko","generated_at":"2026-05-21","decoding":"not-applicable","postprocess":{"editing_pass":"none"},"redistribution":"hash-only","text_hash":"sha256:b60f46512232af2bf08ae38f45c1ca364cf6869b4423a94cbf633fede042db94","source_review":{"status":"template-placeholder","rationale":"Replace text/hash locally before using as evidence; do not commit raw text unless redistribution is cleared."},"class":"natural-human","register":"chat-update","model_family":"human-reference","provider":"local-human-control","model":"maintainer-or-approved-human-source","prompt_id":"rb25-ko-chat-update-001","reviewer_notes":"Native Korean control candidate; keep raw text local unless explicit fixture permission exists."}
+{"sample_id":"ko-ai-01","language":"ko","generated_at":"2026-05-21","decoding":{"temperature":0.7,"top_p":0.9},"postprocess":{"editing_pass":"none"},"redistribution":"hash-only","text_hash":"sha256:e7fb5c91d8e41e685afd24fd43be0ae329c625316b4362614461ad9d97cffe64","source_review":{"status":"template-placeholder","rationale":"Replace text/hash locally before using as evidence; do not commit raw text unless redistribution is cleared."},"class":"ai-like","register":"blog","model_family":"gpt-family","provider":"openai","model":"replace-with-current-gpt-model","prompt_id":"rb25-ko-blog-001","reviewer_notes":"Generated from repo-owned prompt; confirm provider redistribution terms before committing text."}
+{"sample_id":"ko-ai-02","language":"ko","generated_at":"2026-05-21","decoding":{"temperature":0.7,"top_p":0.9},"postprocess":{"editing_pass":"none"},"redistribution":"hash-only","text_hash":"sha256:8cb64dabc655a9b8c1259ef928169ddb44cb5804c1450970688a162a6ee34fa2","source_review":{"status":"template-placeholder","rationale":"Replace text/hash locally before using as evidence; do not commit raw text unless redistribution is cleared."},"class":"ai-like","register":"product-doc","model_family":"gpt-family","provider":"openai","model":"replace-with-current-gpt-model","prompt_id":"rb25-ko-product-doc-001","reviewer_notes":"Generated from repo-owned prompt; confirm provider redistribution terms before committing text."}
+{"sample_id":"ko-ai-03","language":"ko","generated_at":"2026-05-21","decoding":{"temperature":0.7,"top_p":0.9},"postprocess":{"editing_pass":"none"},"redistribution":"hash-only","text_hash":"sha256:aa79ea5db563218758b675447eecfc08b0802aeb986ef8a52ae36ccd637643c7","source_review":{"status":"template-placeholder","rationale":"Replace text/hash locally before using as evidence; do not commit raw text unless redistribution is cleared."},"class":"ai-like","register":"blog","model_family":"claude-family","provider":"anthropic","model":"replace-with-current-claude-model","prompt_id":"rb25-ko-blog-001","reviewer_notes":"Generated from repo-owned prompt; confirm provider redistribution terms before committing text."}
+{"sample_id":"ko-ai-04","language":"ko","generated_at":"2026-05-21","decoding":{"temperature":0.7,"top_p":0.9},"postprocess":{"editing_pass":"none"},"redistribution":"hash-only","text_hash":"sha256:d913797c36a6d0cbcaf3b73e4f05bd2bcf193447f36538072ef86fdb8c51d79c","source_review":{"status":"template-placeholder","rationale":"Replace text/hash locally before using as evidence; do not commit raw text unless redistribution is cleared."},"class":"ai-like","register":"product-doc","model_family":"claude-family","provider":"anthropic","model":"replace-with-current-claude-model","prompt_id":"rb25-ko-product-doc-001","reviewer_notes":"Generated from repo-owned prompt; confirm provider redistribution terms before committing text."}
+{"sample_id":"ko-ai-05","language":"ko","generated_at":"2026-05-21","decoding":{"temperature":0.7,"top_p":0.9},"postprocess":{"editing_pass":"none"},"redistribution":"hash-only","text_hash":"sha256:f3f9caeb75ea6cbb01bc69fd9b54f88fe4205eb8423e5892526f141467a1b16b","source_review":{"status":"template-placeholder","rationale":"Replace text/hash locally before using as evidence; do not commit raw text unless redistribution is cleared."},"class":"ai-like","register":"blog","model_family":"gemini-family","provider":"google","model":"replace-with-current-gemini-model","prompt_id":"rb25-ko-blog-001","reviewer_notes":"Generated from repo-owned prompt; confirm provider redistribution terms before committing text."}
+{"sample_id":"ko-ai-06","language":"ko","generated_at":"2026-05-21","decoding":{"temperature":0.7,"top_p":0.9},"postprocess":{"editing_pass":"none"},"redistribution":"hash-only","text_hash":"sha256:f2a6d0b0f419572aa3e026a3128b349e80246bdb983220838aeb0791c35e477d","source_review":{"status":"template-placeholder","rationale":"Replace text/hash locally before using as evidence; do not commit raw text unless redistribution is cleared."},"class":"ai-like","register":"product-doc","model_family":"gemini-family","provider":"google","model":"replace-with-current-gemini-model","prompt_id":"rb25-ko-product-doc-001","reviewer_notes":"Generated from repo-owned prompt; confirm provider redistribution terms before committing text."}
+{"sample_id":"ko-ai-07","language":"ko","generated_at":"2026-05-21","decoding":{"temperature":0.7,"top_p":0.9},"postprocess":{"editing_pass":"none"},"redistribution":"hash-only","text_hash":"sha256:6fc0a933e288ab1947152b67403a0b43848b9237a8c0e6fc6c8923913c181569","source_review":{"status":"template-placeholder","rationale":"Replace text/hash locally before using as evidence; do not commit raw text unless redistribution is cleared."},"class":"ai-like","register":"blog","model_family":"open-weight","provider":"local-or-hosted-open-weight","model":"replace-with-open-weight-model","prompt_id":"rb25-ko-blog-001","reviewer_notes":"Generated from repo-owned prompt; confirm provider redistribution terms before committing text."}
+{"sample_id":"ko-ai-08","language":"ko","generated_at":"2026-05-21","decoding":{"temperature":0.7,"top_p":0.9},"postprocess":{"editing_pass":"none"},"redistribution":"hash-only","text_hash":"sha256:48570564c3672a6e02cf459fc9adb4c7bf1c5f159f70d11dcb02e0e9f96627e7","source_review":{"status":"template-placeholder","rationale":"Replace text/hash locally before using as evidence; do not commit raw text unless redistribution is cleared."},"class":"ai-like","register":"product-doc","model_family":"open-weight","provider":"local-or-hosted-open-weight","model":"replace-with-open-weight-model","prompt_id":"rb25-ko-product-doc-001","reviewer_notes":"Generated from repo-owned prompt; confirm provider redistribution terms before committing text."}
+{"sample_id":"ko-lightedit-01","language":"ko","generated_at":"2026-05-21","decoding":{"temperature":0.7,"top_p":0.9},"postprocess":{"editing_pass":"light human edit"},"redistribution":"hash-only","text_hash":"sha256:1bd8d748521cd2abf1f164b38f0676f9596e729f9a6d5ff6b0cd0d113bfd3937","source_review":{"status":"template-placeholder","rationale":"Replace text/hash locally before using as evidence; do not commit raw text unless redistribution is cleared."},"class":"lightly-edited-ai","register":"product-doc","model_family":"gpt-family","provider":"openai","model":"replace-with-gpt-family-model","prompt_id":"rb25-ko-edit-light-001","reviewer_notes":"Store before/after hashes locally; do not use for public claims until paired evidence is complete."}
+{"sample_id":"ko-lightedit-02","language":"ko","generated_at":"2026-05-21","decoding":{"temperature":0.7,"top_p":0.9},"postprocess":{"editing_pass":"light human edit"},"redistribution":"hash-only","text_hash":"sha256:e8b8aaa4f8570df6d89019f84bb0fea043b580907e63f0c28011c105ed307d64","source_review":{"status":"template-placeholder","rationale":"Replace text/hash locally before using as evidence; do not commit raw text unless redistribution is cleared."},"class":"lightly-edited-ai","register":"chat-update","model_family":"claude-family","provider":"anthropic","model":"replace-with-claude-family-model","prompt_id":"rb25-ko-edit-light-001","reviewer_notes":"Store before/after hashes locally; do not use for public claims until paired evidence is complete."}
+{"sample_id":"ko-heavyedit-01","language":"ko","generated_at":"2026-05-21","decoding":{"temperature":0.7,"top_p":0.9},"postprocess":{"editing_pass":"heavy human edit"},"redistribution":"hash-only","text_hash":"sha256:f9eece5ccfadfb883fce879ceddad5a2cfa04b636e8712c1c4d9036e34d7adc4","source_review":{"status":"template-placeholder","rationale":"Replace text/hash locally before using as evidence; do not commit raw text unless redistribution is cleared."},"class":"heavily-edited-ai","register":"blog","model_family":"gemini-family","provider":"google","model":"replace-with-gemini-family-model","prompt_id":"rb25-ko-edit-heavy-001","reviewer_notes":"Store before/after hashes locally; do not use for public claims until paired evidence is complete."}
+{"sample_id":"ko-heavyedit-02","language":"ko","generated_at":"2026-05-21","decoding":{"temperature":0.7,"top_p":0.9},"postprocess":{"editing_pass":"heavy human edit"},"redistribution":"hash-only","text_hash":"sha256:dde36e9ecaf2b770f6180012477f0cdf124be010fe0ace86f2e5e333577022d3","source_review":{"status":"template-placeholder","rationale":"Replace text/hash locally before using as evidence; do not commit raw text unless redistribution is cleared."},"class":"heavily-edited-ai","register":"technical-how-to","model_family":"open-weight","provider":"local-or-hosted-open-weight","model":"replace-with-open-weight-model","prompt_id":"rb25-ko-edit-heavy-001","reviewer_notes":"Store before/after hashes locally; do not use for public claims until paired evidence is complete."}
+{"sample_id":"ko-katfish-essay-01","language":"ko","generated_at":"2026-05-21","decoding":"not-applicable","postprocess":{"editing_pass":"none"},"redistribution":"hash-only","text_hash":"sha256:4002d6936d149b24285e9c2965c5b5c567f9a782e34588ebe24812eeed2606cb","source_review":{"status":"license-review","rationale":"Use hash-only until KatFish redistribution terms are recorded for this repo."},"class":"ai-like","register":"academic-summary","model_family":"open-weight","provider":"katfishnet","model":"katfish-essay-source","prompt_id":"katfish-essay-row-id-replace-me","source_url":"https://github.com/Shinwoo-Park/katfishnet","reviewer_notes":"KatFish essay comparison row; replace hash with local row digest."}
+{"sample_id":"ko-katfish-poetry-01","language":"ko","generated_at":"2026-05-21","decoding":"not-applicable","postprocess":{"editing_pass":"none"},"redistribution":"hash-only","text_hash":"sha256:fcdd62c1d162426742044a7334341d9513cee1eb657b2eb4a37d500224ac3982","source_review":{"status":"license-review","rationale":"Use hash-only until KatFish redistribution terms are recorded for this repo."},"class":"ai-like","register":"blog","model_family":"open-weight","provider":"katfishnet","model":"katfish-poetry-source","prompt_id":"katfish-poetry-row-id-replace-me","source_url":"https://github.com/Shinwoo-Park/katfishnet","reviewer_notes":"KatFish poetry comparison row; replace hash with local row digest."}
+{"sample_id":"ko-katfish-abstract-01","language":"ko","generated_at":"2026-05-21","decoding":"not-applicable","postprocess":{"editing_pass":"none"},"redistribution":"hash-only","text_hash":"sha256:70cdc34465d20b4c18a623fb222e5382a7807c690f534bb593ad52ffcba03d28","source_review":{"status":"license-review","rationale":"Use hash-only until KatFish redistribution terms are recorded for this repo."},"class":"ai-like","register":"academic-summary","model_family":"open-weight","provider":"katfishnet","model":"katfish-paper-abstract-source","prompt_id":"katfish-paper-abstract-row-id-replace-me","source_url":"https://github.com/Shinwoo-Park/katfishnet","reviewer_notes":"KatFish paper-abstract comparison row; replace hash with local row digest."}
+{"sample_id":"ko-fp-learner-01","language":"ko","generated_at":"2026-05-21","decoding":"not-applicable","postprocess":{"editing_pass":"none"},"redistribution":"hash-only","text_hash":"sha256:74c4bb285a763e1de64f8490bdc0e51299b6b02b5d636f7c53d7c169f96e844a","source_review":{"status":"hash-only","rationale":"Learner-writing stress row mapped to academic-summary until register schema expands."},"class":"natural-human","register":"academic-summary","model_family":"human-reference","provider":"learner-corpus","model":"human-reference","prompt_id":"learner-corpus-replace-me","reviewer_notes":"Learner-writing stress row mapped to academic-summary until register schema expands."}
+{"sample_id":"ko-fp-community-01","language":"ko","generated_at":"2026-05-21","decoding":"not-applicable","postprocess":{"editing_pass":"none"},"redistribution":"hash-only","text_hash":"sha256:a8c399142311d495e36904d801875afbb7d8d51fef9c350813376f2414ef02e5","source_review":{"status":"hash-only","rationale":"Community false-positive row; use only with explicit fixture permission."},"class":"natural-human","register":"chat-update","model_family":"human-reference","provider":"false-positive-form","model":"human-reference","prompt_id":"false-positive-form-replace-me","reviewer_notes":"Community false-positive row; use only with explicit fixture permission."}

package/artifacts/rebaseline-2025/prompts.template.jsonl

@@ -0,0 +1,7 @@
+{"prompt_id":"rb25-ko-academic-da-001","language":"ko","register":"academic-summary","purpose":"native-human-control or model generation anchor","prompt":"한국어 학술 요약체(종결-다)로, 특정 연구 결과를 120~180자 한 문단으로 요약한다. 과장된 표현 없이 방법, 결과, 한계를 모두 포함한다."}
+{"prompt_id":"rb25-ko-product-doc-001","language":"ko","register":"product-doc","purpose":"native-human-control or model generation anchor","prompt":"개발자 문서의 기능 설명처럼 120~180자 한 문단을 작성한다. 설정값, 예외 조건, 사용자가 확인할 항목을 구체적으로 쓴다."}
+{"prompt_id":"rb25-ko-blog-001","language":"ko","register":"blog","purpose":"model generation anchor","prompt":"개인 블로그 톤으로 도구를 써본 짧은 후기를 120~180자 한 문단으로 작성한다. 장점 하나와 아쉬운 점 하나를 모두 포함한다."}
+{"prompt_id":"rb25-ko-chat-update-001","language":"ko","register":"chat-update","purpose":"model generation anchor","prompt":"팀 채팅 업데이트처럼 오늘 진행한 일, 막힌 점, 다음 액션을 100~160자 한 문단으로 정리한다."}
+{"prompt_id":"rb25-ko-technical-howto-001","language":"ko","register":"technical-how-to","purpose":"model generation anchor","prompt":"터미널 명령을 실행하기 전 확인해야 할 조건과 실패 시 되돌리는 방법을 120~180자 한 문단으로 설명한다."}
+{"prompt_id":"rb25-ko-edit-light-001","language":"ko","register":"product-doc","purpose":"lightly-edited-ai","prompt":"AI가 작성한 제품 문서 문단을 사람이 맞춤법과 어색한 조사만 살짝 고친 버전으로 만든다. 의미와 문장 구조는 대부분 유지한다."}
+{"prompt_id":"rb25-ko-edit-heavy-001","language":"ko","register":"blog","purpose":"heavily-edited-ai","prompt":"AI가 작성한 블로그 문단을 사람이 구조와 어휘를 크게 바꾼 버전으로 만든다. 핵심 주장과 사실은 유지한다."}

package/artifacts/rebaseline-2025/sources.ko-public.jsonl

@@ -0,0 +1,39 @@
+{"source_id":"korea-policy-sme-2025","url":"https://www.korea.kr/news/policyNewsView.do?newsId=148938345","register":"chat-update","source_title":"소상공인 전환보증 3조 2000억 공급…전국 30곳 채무조정센터 설치","source_license":"Public web page; news images/third-party media excluded; raw text remains hash-only pending page-level redistribution review.","source_published_at":"2025-01-08","max_rows":10}
+{"source_id":"korea-policy-export-2025","url":"https://www.korea.kr/news/policyNewsView.do?newsId=148948589","register":"chat-update","source_title":"미 상호관세 충격 수출 중소기업 지원 3대 지원책 가동","source_license":"Public web page; news images/third-party media excluded; raw text remains hash-only pending page-level redistribution review.","source_published_at":"2025-09-03","max_rows":10}
+{"source_id":"korea-policy-budget-2025","url":"https://www.korea.kr/news/policyNewsView.do?newsId=148942874","register":"chat-update","source_title":"2025년 제1차 추가경정예산 4.8조 원 확정","source_license":"Public web page; news images/third-party media excluded; raw text remains hash-only pending page-level redistribution review.","source_published_at":"2025-05-02","max_rows":10}
+{"source_id":"korea-policy-education-2025","url":"https://www.korea.kr/news/policyNewsView.do?newsId=148939202","register":"chat-update","source_title":"주거안정장학금 첫 시행…대학생에 월 최대 20만 원 지원","source_license":"Public web page; news images/third-party media excluded; raw text remains hash-only pending page-level redistribution review.","source_published_at":"2025-02-04","max_rows":10}
+{"source_id":"korea-policy-car-2025","url":"https://www.korea.kr/news/policyNewsView.do?newsId=148941486","register":"chat-update","source_title":"자동차·부품 업계에 정책금융 2조 원 추가 투입","source_license":"Public web page; news images/third-party media excluded; raw text remains hash-only pending page-level redistribution review.","source_published_at":"2025-04-09","max_rows":10}
+{"source_id":"toss-navigation-score-2025","url":"https://toss.tech/article/Toss_Navigation_Score","register":"blog","source_title":"토스가 특허 낸 리서치툴, TNS 제작기","source_license":"Public web page; no redistribution review complete; raw text remains hash-only.","source_published_at":"2025-07-04","max_rows":10}
+{"source_id":"toss-bank-interns-2025","url":"https://toss.tech/article/toss-bank-interns","register":"blog","source_title":"슬기로운 토스뱅크 개발 인턴 생활","source_license":"Public web page; no redistribution review complete; raw text remains hash-only.","source_published_at":"2025-07-14","max_rows":10}
+{"source_id":"toss-frontend-github-2025","url":"https://toss.tech/article/34895","register":"blog","source_title":"토스 프론트엔드에 이력서 없이 리포지토리 링크로 지원하세요","source_license":"Public web page; no redistribution review complete; raw text remains hash-only.","source_published_at":"2025-04-08","max_rows":10}
+{"source_id":"toss-people-design-2025","url":"https://toss.tech/article/Tosspeople_LeeJiyoon","register":"blog","source_title":"토스 피플: 이것도 나니까 할 수 있다고 생각하기","source_license":"Public web page; no redistribution review complete; raw text remains hash-only.","source_published_at":"2025-03-27","max_rows":10}
+{"source_id":"toss-ml-platform-2025","url":"https://toss.tech/article/feature-store-trainkit","register":"technical-how-to","source_title":"토스가 다양한 ML 모델을 만드는 법: Feature Store & Trainkit","source_license":"Public web page; no redistribution review complete; raw text remains hash-only.","source_published_at":"2025-08-14","max_rows":12}
+{"source_id":"toss-gpu-mig-2025","url":"https://toss.tech/article/toss-securities-gpu-mig","register":"technical-how-to","source_title":"GPU를 밀도 있게 쓰는 방법 - 토스증권의 GPU 가상화(MIG) 도입기","source_license":"Public web page; no redistribution review complete; raw text remains hash-only.","source_published_at":"2025-07-10","max_rows":12}
+{"source_id":"kistep-rnd-agenda-2025","url":"https://www.kistep.re.kr/reportDetail.es?mid=a10305050000&rpt_no=RES0220260011&rpt_tp=831-005","register":"academic-summary","source_title":"2025년 국가연구개발 투자 핵심이슈 발굴 및 아젠다 수립 연구","source_license":"Public research summary page; raw text remains hash-only pending redistribution review.","source_published_at":"2026-02-11","max_rows":10}
+{"source_id":"kei-sdgs-2025","url":"https://repository.kei.re.kr/handle/2017.oak/24270","register":"academic-summary","source_title":"지속가능발전목표(SDGs)의 상호작용에 관한 연구","source_license":"Public repository abstract page; raw text remains hash-only pending redistribution review.","source_published_at":"2025-03-31","max_rows":8}
+{"source_id":"nrf-trend-ai-2025","url":"https://webzine.nrf.re.kr/magazine/2503/sub3.php","register":"academic-summary","source_title":"한국연구재단웹진 2025년 3월 트렌드 리포트","source_license":"Public webzine page; raw text remains hash-only pending redistribution review.","source_published_at":"2025-03-01","max_rows":8}
+{"source_id":"kakao-login-doc","url":"https://developers.kakao.com/docs/latest/ko/kakaologin/common","register":"product-doc","source_title":"카카오 로그인 공통 가이드","source_license":"Public developer documentation; no redistribution review complete; raw text remains hash-only.","max_rows":12}
+{"source_id":"naver-login-doc","url":"https://developers.naver.com/docs/login/overview/overview.md","register":"product-doc","source_title":"네이버 로그인 API 명세","source_license":"Public developer documentation; no redistribution review complete; raw text remains hash-only.","max_rows":12}
+{"source_id":"seoul-opengov-copyright","url":"https://opengov.seoul.go.kr/copyright","register":"product-doc","source_title":"서울 열린데이터광장 저작권 정책","source_license":"Public-sector copyright policy page; raw text remains hash-only until row review.","max_rows":10}
+{"source_id":"korea-policy-lowbirth-2025","url":"https://www.korea.kr/news/policyNewsView.do?newsId=148945476","register":"chat-update","source_title":"올해 저출산·고령화 대응에 100조 7000억 원 투입…성과기반 운영","source_license":"Public web page; news images/third-party media excluded; raw text remains hash-only pending page-level redistribution review.","source_published_at":"2025-07-04","max_rows":10}
+{"source_id":"korea-policy-youth-work-2025","url":"https://www.korea.kr/news/policyNewsView.do?newsId=148947124","register":"chat-update","source_title":"청년 근속 인센티브 조기 지급…7월부터 청년 3282명에","source_license":"Public web page; news images/third-party media excluded; raw text remains hash-only pending page-level redistribution review.","source_published_at":"2025-07-01","max_rows":6}
+{"source_id":"toss-front-2","url":"https://toss.tech/article/toss_front","register":"blog","source_title":"하마터면 못생겨질 뻔했다 - 토스 프론트 2 제작기","source_license":"Public web page; no redistribution review complete; raw text remains hash-only.","max_rows":10}
+{"source_id":"toss-tpm-ai-org","url":"https://toss.tech/article/toss-tpm","register":"blog","source_title":"AI 시대, 성과 내는 조직일수록 토스식 TPM이 필요한 이유","source_license":"Public web page; no redistribution review complete; raw text remains hash-only.","max_rows":10}
+{"source_id":"toss-da-assistant-panda","url":"https://toss.tech/article/da-assistant-panda","register":"technical-how-to","source_title":"토스플레이스 데이터봇 판다(PANDA)를 소개합니다","source_license":"Public web page; no redistribution review complete; raw text remains hash-only.","max_rows":12}
+{"source_id":"toss-starrocks-resource-group","url":"https://toss.tech/article/operating-starrocks-1","register":"technical-how-to","source_title":"StarRocks 운영기: Resource Group으로 멀티테넌트 워크로드 격리하기","source_license":"Public web page; no redistribution review complete; raw text remains hash-only.","max_rows":12}
+{"source_id":"toss-post-quantum-crypto","url":"https://toss.tech/article/post-quantum-cryptography","register":"technical-how-to","source_title":"양자컴퓨터 시대에 대비한 양자내성암호 적용","source_license":"Public web page; no redistribution review complete; raw text remains hash-only.","max_rows":12}
+{"source_id":"toss-harness-productivity","url":"https://toss.tech/article/harness-for-team-productivity","register":"technical-how-to","source_title":"Software 3.0 시대, Harness를 통한 조직 생산성 저점 높이기","source_license":"Public web page; no redistribution review complete; raw text remains hash-only.","max_rows":12}
+{"source_id":"kakao-login-faq","url":"https://developers.kakao.com/docs/latest/ko/kakaologin/faq","register":"product-doc","source_title":"카카오 로그인 FAQ","source_license":"Public developer documentation; no redistribution review complete; raw text remains hash-only.","max_rows":12}
+{"source_id":"kakao-login-flutter","url":"https://developers.kakao.com/docs/latest/ko/kakaologin/flutter","register":"product-doc","source_title":"카카오 로그인 Flutter 가이드","source_license":"Public developer documentation; no redistribution review complete; raw text remains hash-only.","max_rows":12}
+{"source_id":"naver-login-api","url":"https://developers.naver.com/docs/login/api/api.md","register":"product-doc","source_title":"네이버 로그인 API 명세","source_license":"Public developer documentation; no redistribution review complete; raw text remains hash-only.","max_rows":10}
+{"source_id":"kakao-android-start","url":"https://developers.kakao.com/docs/latest/ko/android/getting-started","register":"product-doc","source_title":"Kakao SDK for Android 시작하기","source_license":"Public developer documentation; no redistribution review complete; raw text remains hash-only.","max_rows":8}
+{"source_id":"kakao-dev-quota","url":"https://developers.kakao.com/docs/latest/ko/getting-started/quota","register":"product-doc","source_title":"Kakao Developers 쿼터","source_license":"Public developer documentation; no redistribution review complete; raw text remains hash-only.","max_rows":4}
+{"source_id":"kistep-performance-eval-2025","url":"https://www.kistep.re.kr/reportAllDetail.es?mid=a10305010000&rpt_no=RES0220260060&rpt_tp=831-001","register":"academic-summary","source_title":"2025년 성과평가 정책수립·운영 및 평가실시","source_license":"Public research summary page; raw text remains hash-only pending redistribution review.","source_published_at":"2026-02-25","max_rows":24}
+{"source_id":"kistep-rnd-consistency-2025","url":"https://www.kistep.re.kr/reportAllDetail.es?mid=a10305010000&rpt_no=RES0220260037&rpt_tp=831-003","register":"academic-summary","source_title":"2025년 R&D사업 예비타당성조사 일관성 제고를 위한 조사체계 개선방향 연구","source_license":"Public research summary page; raw text remains hash-only pending redistribution review.","source_published_at":"2026-02-08","max_rows":8}
+{"source_id":"kistep-outcome-policy-2025","url":"https://www.kistep.re.kr/reportDetail.es?mid=a10305050000&rpt_no=RES0220260095&rpt_tp=831-004","register":"academic-summary","source_title":"2025년 연구성과 확산 정책 수립 및 제도 운영","source_license":"Public research summary page; raw text remains hash-only pending redistribution review.","source_published_at":"2026-03-30","max_rows":8}
+{"source_id":"kistep-rnd-investment-issue-2025","url":"https://www.kistep.re.kr/reportDetail.es?mid=a10305050000&rpt_no=RES0220260070&rpt_tp=831-004","register":"academic-summary","source_title":"2025년 R&D 투자의 성과 활성화 정책이슈 분석 연구","source_license":"Public research summary page; raw text remains hash-only pending redistribution review.","source_published_at":"2026-02-25","max_rows":8}
+{"source_id":"kistep-scoreboard-2025","url":"https://www.kistep.re.kr/reportAllDetail.es?mid=a10305010000&rpt_no=RES0220260028&rpt_tp=831-002","register":"academic-summary","source_title":"2025년 과학기술혁신정책 스코어보드 개발 연구","source_license":"Public research summary page; raw text remains hash-only pending redistribution review.","source_published_at":"2026-02-25","max_rows":8}
+{"source_id":"korea-policy-youth-basic-2025","url":"https://www.korea.kr/news/policyNewsView.do?newsId=148949763","register":"chat-update","source_title":"정부, 보편적 청년정책 추진…안정적 기본생활 보장","source_license":"Public web page; news images/third-party media excluded; raw text remains hash-only pending page-level redistribution review.","source_published_at":"2025-09-22","max_rows":6}
+{"source_id":"kistep-student-researcher-2025","url":"https://www.kistep.re.kr/reportDetail.es?mid=a10305090000&rpt_no=RES0220260096&rpt_tp=831-002","register":"academic-summary","source_title":"2025년 학생인건비 및 학생연구자 지원 제도 운영 연구","source_license":"Public research summary page; raw text remains hash-only pending redistribution review.","source_published_at":"2026-03-31","max_rows":4}
+{"source_id":"kistep-research-ethics-2025","url":"https://www.kistep.re.kr/reportAllDetail.es?mid=a10305010000&rpt_no=RES0220260054&rpt_tp=831-004","register":"academic-summary","source_title":"2025년 연구윤리 확보 및 실무지원 연구","source_license":"Public research summary page; raw text remains hash-only pending redistribution review.","source_published_at":"2026-02-20","max_rows":4}
+{"source_id":"kistep-preliminary-feasibility-2025","url":"https://www.kistep.re.kr/reportAllDetail.es?mid=a10305010000&rpt_no=RES0220260093&rpt_tp=831-003","register":"academic-summary","source_title":"2025년 국가연구개발사업 예비타당성조사 제도 운영 연구","source_license":"Public research summary page; raw text remains hash-only pending redistribution review.","source_published_at":"2026-03-31","max_rows":4}

package/assets/brand/patina-badge.svg

@@ -0,0 +1,18 @@
+<svg xmlns="http://www.w3.org/2000/svg" width="164" height="28" viewBox="0 0 164 28" role="img" aria-labelledby="title desc">
+  <title id="title">patina score badge</title>
+  <desc id="desc">Static patina README badge for repositories that want the mark without publishing a live score endpoint.</desc>
+  <defs>
+    <linearGradient id="score" x1="0" y1="0" x2="1" y2="1">
+      <stop offset="0%" stop-color="#2dd4bf"/>
+      <stop offset="100%" stop-color="#34d399"/>
+    </linearGradient>
+  </defs>
+  <rect width="164" height="28" rx="14" fill="#020617"/>
+  <rect x="1" y="1" width="162" height="26" rx="13" fill="none" stroke="#1e293b"/>
+  <path d="M13 11c3.2-5.4 11-6 15.8-1.8l-3.5 3.5c-2.8-2.2-6.2-1.8-8.2 1z" fill="#c46a2a"/>
+  <path d="M28.8 17c-3.2 5.4-11 6-15.8 1.8l3.3-3.7c3 2.5 6.4 2.1 8.5-.8z" fill="#2dd4bf"/>
+  <circle cx="20.8" cy="14" r="2.7" fill="#ffe6a8"/>
+  <text x="38" y="18" fill="#f8fafc" font-family="Inter, ui-sans-serif, system-ui, -apple-system, BlinkMacSystemFont, Segoe UI, sans-serif" font-size="12" font-weight="800" letter-spacing="-.2">patina</text>
+  <rect x="83" y="4" width="77" height="20" rx="10" fill="url(#score)"/>
+  <text x="93" y="18" fill="#042f2e" font-family="Inter, ui-sans-serif, system-ui, -apple-system, BlinkMacSystemFont, Segoe UI, sans-serif" font-size="11" font-weight="800">human-ish</text>
+</svg>

package/assets/brand/patina-mark.svg

@@ -0,0 +1,8 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 512 512" role="img" aria-labelledby="title desc">
+  <title id="title">patina mark</title>
+  <desc id="desc">A transparent copper-to-teal mark around a warm preserved-meaning core.</desc>
+
+  <path d="M92 196C160 86 320 74 420 164L346 238C288 190 218 198 174 258Z" fill="#c46a2a"/>
+  <path d="M420 316C352 426 192 438 92 348L160 270C224 322 294 314 338 254Z" fill="#2dd4bf"/>
+  <circle cx="252" cy="248" r="54" fill="#ffe6a8" stroke="#020617" stroke-width="10"/>
+</svg>

package/assets/demo/README.md

@@ -0,0 +1,79 @@
+# Demo assets
+
+README hero animations are language-suffixed so each localized README can point
+at a matching demo instead of reusing a Korean recording everywhere.
+
+Current assets:
+
+- `patina-demo-en.gif` — English README hero. `README_ZH.md` and `README_JA.md`
+  intentionally fall back to this English terminal demo until localized ZH/JA
+  recordings are worth maintaining.
+  - source: `examples/short/marketing-launch-en.md`
+  - expected rewrite: `examples/short/marketing-launch-en-rewritten.md`
+- `patina-demo-ko.gif` — Korean README hero.
+  - source: `examples/short/marketing-launch.md`
+  - expected rewrite: `examples/short/marketing-launch-rewritten.md`
+- Future localized demos should use the same naming pattern:
+  `patina-demo-zh.gif`, `patina-demo-ja.gif`.
+
+Shared requirements:
+
+- verification: each rewritten fixture should pass the 30% content gate before
+  the GIF is updated
+- size target: GIF under 10 MB so GitHub renders it reliably
+
+## Deterministic fallback render
+
+When a live terminal recorder is not available, render a small transcript GIF
+from checked-in fixtures. This helper is optional and requires local Python +
+Pillow only for asset regeneration; it is not a package runtime dependency.
+
+```bash
+python3 scripts/render-demo-gif.py \
+  --lang en \
+  --source examples/short/marketing-launch-en.md \
+  --rewrite examples/short/marketing-launch-en-rewritten.md \
+  --output assets/demo/patina-demo-en.gif \
+  --title "patina demo — English" \
+  --score-line "PASS · score 0.0% · MPS: meaning preserved"
+
+node scripts/precommit-score.mjs examples/short/marketing-launch-en-rewritten.md
+ls -lh assets/demo/patina-demo-en.gif
+
+# Korean variant, when it needs to be regenerated:
+python3 scripts/render-demo-gif.py \
+  --lang ko \
+  --source examples/short/marketing-launch.md \
+  --rewrite examples/short/marketing-launch-rewritten.md \
+  --output assets/demo/patina-demo-ko.gif \
+  --title "patina demo — Korean" \
+  --score-line "PASS · score under 30% · MPS: meaning preserved"
+
+node scripts/precommit-score.mjs examples/short/marketing-launch-rewritten.md
+```
+
+## Preferred live terminal capture
+
+Use this when `asciinema` and `agg` are installed and you want to capture a real
+terminal session instead of the deterministic fallback render:
+
+```bash
+# 1) Record a real terminal session using the checked-in language fixture.
+asciinema rec /tmp/patina-demo-en.cast
+
+# In the recording, run:
+cat examples/short/marketing-launch-en.md
+patina --lang en --tone marketing examples/short/marketing-launch-en.md
+node scripts/precommit-score.mjs examples/short/marketing-launch-en-rewritten.md
+
+# 2) Render to a GitHub-safe GIF. Do not use animated SVG for README motion.
+agg /tmp/patina-demo-en.cast assets/demo/patina-demo-en.gif \
+  --cols 82 --rows 24 \
+  --font-family "DejaVu Sans Mono"
+
+# 3) Keep the asset small and verify the rewritten fixture still passes.
+ls -lh assets/demo/patina-demo-en.gif
+node scripts/precommit-score.mjs examples/short/marketing-launch-en-rewritten.md
+```
+
+If `agg` output is too large, compress with `gifsicle -O3` or re-record a shorter cast.

package/core/scoring.md

@@ -76,52 +76,52 @@ Unknown categories (from custom packs not in the weight config) get default weig
 | Category | Weight | Patterns | Notes |
 |----------|--------|----------|-------|
 | content | 0.18 | 6 | |
-| language | 0.18 | 7 | |
+| language | 0.18 | 8 | |
 | style | 0.18 | 6 | |
 | communication | 0.13 | 4 | |
 | filler | 0.08 | 4 | |
 | structure | 0.15 | 5 | |
-| viral-hook | 0.10 | 8 | score-only (no rewrite) |
-| **Total** | **1.00** | **40** | |
+| viral-hook | 0.10 | 9 | score-only (no rewrite) |
+| **Total** | **1.00** | **42** | |
 
 ### English (en)
 
 | Category | Weight | Patterns | Notes |
 |----------|--------|----------|-------|
 | content | 0.20 | 6 | |
-| language | 0.20 | 7 | |
+| language | 0.20 | 8 | |
 | style | 0.20 | 6 | |
 | communication | 0.12 | 4 | |
 | filler | 0.08 | 4 | |
 | structure | 0.10 | 5 | |
-| viral-hook | 0.10 | 8 | score-only (no rewrite) |
-| **Total** | **1.00** | **40** | |
+| viral-hook | 0.10 | 9 | score-only (no rewrite) |
+| **Total** | **1.00** | **42** | |
 
 ### Chinese (zh)
 
 | Category | Weight | Patterns | Notes |
 |----------|--------|----------|-------|
 | content | 0.18 | 6 | |
-| language | 0.18 | 7 | |
+| language | 0.18 | 8 | |
 | style | 0.18 | 6 | |
 | communication | 0.13 | 4 | |
 | filler | 0.08 | 4 | |
 | structure | 0.15 | 5 | |
-| viral-hook | 0.10 | 8 | score-only (no rewrite) |
-| **Total** | **1.00** | **40** | |
+| viral-hook | 0.10 | 9 | score-only (no rewrite) |
+| **Total** | **1.00** | **42** | |
 
 ### Japanese (ja)
 
 | Category | Weight | Patterns | Notes |
 |----------|--------|----------|-------|
 | content | 0.18 | 6 | |
-| language | 0.18 | 7 | |
+| language | 0.18 | 8 | |
 | style | 0.18 | 6 | |
 | communication | 0.13 | 4 | |
 | filler | 0.08 | 4 | |
 | structure | 0.15 | 5 | |
-| viral-hook | 0.10 | 8 | score-only (no rewrite) |
-| **Total** | **1.00** | **40** | |
+| viral-hook | 0.10 | 9 | score-only (no rewrite) |
+| **Total** | **1.00** | **42** | |
 
 Weights are configurable via `ouroboros.category-weights.{lang}` in `.patina.yaml`.
 

package/core/standalone-prompt.md

@@ -123,7 +123,7 @@ FOR each anchor IN anchor_list:
 Apply all remaining pattern packs (content, language, style, communication, filler).
 
 1. **AI pattern identification** — scan all loaded sentence/lexical patterns
-2. **Problem segment rewrite** — replace AI-sounding expressions with natural alternatives
+2. **Problem segment rewrite** — do not swap tokens in place; read the local context and rewrite the affected clause/sentence into a natural alternative
 3. **Meaning preservation** — keep core message intact
 4. **Tone matching** — adjust tone per the profile's guidance
 5. **Voice injection** — add personality per `core/voice.md`
@@ -134,6 +134,8 @@ Apply all remaining pattern packs (content, language, style, communication, fill
    - MEDIUM semantic risk: inject only Polarity/Negation anchors
    - LOW semantic risk: no constraints
 
+**CJK clause-level rewrite guard (issue #352):** For `ko`, `zh`, and `ja`, do not fix AI tells by replacing one punctuation mark or one token at a time. If connective punctuation (em dash, colon, semicolon, slash, comma splice, parenthetical aside) appears with a suspect phrase, read the whole sentence and choose an idiomatic clause structure, sentence split, or connective phrase in the target language. If a translationese/calque phrase is attached to punctuation, fix both together at clause level. Korean examples: prefer `TUI 없이 완전 자율로 설치하려면 ...` over `무 TUI ...`, and `"끝난 것 같아요"만으로는 부족한, 결과를 끝까지 확인해야 하는 열린 작업` over `"끝난 것 같아요"로는 부족한 열린 작업`. Preserve actors, polarity, conditions, numbers, and causation.
+
 **Caution**: Do NOT re-tidy sections already corrected in Phase 1 back into "polished officialese".
 
 #### 5b-v: Anchor Verification

package/core/stylometry.md

@@ -24,6 +24,8 @@ description: Deterministic statistical preprocessing that flags suspect paragrap
   hot 단락/문장 정보를 5a·5b 입력에 주입한다
 - 외부 의존성(형태소 분석기, 외부 detector API) 없이 ko/en은 whitespace 토큰화,
   zh/ja는 deterministic character-token fallback으로 동작한다
+- ko는 dependency-free 보조 지표(띄어쓰기 규칙성, 쉼표 밀도, 조사/어미 suffix diversity
+  proxy)를 함께 기록하고, 세 지표가 모두 보수적 임계값을 넘을 때만 ko 전용 hot 신호로 사용한다
 
 이 알고리즘은 5a 구조 분석 단계와 5b 문장/어휘 단계에 자연스럽게 결합되도록 설계되었다.
 - 5a는 단락 수준 hot 표시를 우선 검토 대상으로 사용한다
@@ -33,7 +35,7 @@ description: Deterministic statistical preprocessing that flags suspect paragrap
 
 | 언어 | 지원 | 비고 |
 |------|---------|------|
-| ko | yes | 어절 단위 토큰화 |
+| ko | yes | 어절 단위 토큰화 + dependency-free diagnostic signals |
 | en | yes | 단어 단위 토큰화 |
 | zh | yes | Han character-token fallback |
 | ja | yes | Kana/Han character-token fallback; 형태소 분석 없음 |
@@ -81,6 +83,10 @@ ELSE:
 > sudachi, mecab 같은 형태소 분석기는 배포 의존성과 설치 실패면을 늘리므로 기본 경로에서
 > 제외한다.
 
+ko diagnostic POS diversity는 형태소 분석기가 아니라 suffix proxy다. 조사/어미 후보 목록으로
+어절 끝을 보수적으로 분류하고, class diversity와 coverage를 기록한다. 이 proxy는 진짜 POS
+tagger가 아니므로 단독 hot 신호가 아니라 spacing/comma와 결합된 ko composite에서만 사용한다.
+
 ---
 
 ## 3. Sentence Splitting
@@ -126,6 +132,8 @@ burstiness_CV = stddev / mean
 - `CV` = Coefficient of Variation (변동계수)
 - mean 이 0 이면 정의되지 않음 → skip 처리
 - population stddev 사용 (표본 보정 없음, 단순화)
+- 단락 문장이 3개 미만이면 CV 값은 기록하되 burstiness band는 부여하지 않는다.
+  2문장 이하 샘플은 길이 변동성이 너무 불안정해서 hot 판정에서 제외한다.
 
 ### 밴드
 
@@ -181,26 +189,68 @@ ELSE:
 
 ---
 
+## 5.1 Korean Diagnostic Signals
+
+한국어는 어절 토큰화만으로는 조사/어미 변형과 띄어쓰기 습관을 충분히 보지 못한다. 그래서
+`analyzeText(text, { lang: "ko" })`는 다음 보조 지표를 단락 결과에 추가한다.
+
+| 필드 | 지표 | 목적 | 현재 판정 영향 |
+|------|------|------|----------------|
+| `spacing` | `eojeolLengthCV`, 평균 어절 길이, 1음절/긴 어절 비율 | LLM식으로 지나치게 균질한 띄어쓰기·어절 길이 후보 추적 | composite 조건 |
+| `comma` | 쉼표 수, 문장당 쉼표, 100자당 쉼표 | 쉼표 리듬이 사라진 장문 후보 추적 | composite 조건 |
+| `posDiversity` | 조사/어미 suffix class coverage/diversity proxy | 형태소 분석기 없이 기능어 다양성 후보 추적 | composite 조건 |
+| `koDiagnostics` | 세 지표의 보수적 AND 판정 | burstiness/MATTR/lexicon이 놓치는 ko hotspot 보강 | hot OR-rule 참여 |
+
+POS proxy 결정:
+- 새 런타임 의존성은 추가하지 않는다.
+- KoNLPy/mecab-ko/khaiii 같은 형태소 분석기는 설치 실패면과 배포 크기가 커서 기본 경로에서
+  제외한다.
+- suffix proxy는 `은/는`, `이/가`, `을/를`, `에게/에서`, `습니다/합니다` 같은 조사·어미
+  class를 세고, `proxy: "suffix"`로 명시한다.
+
+운영 규칙:
+- 세 조건을 모두 만족할 때만 `koDiagnostics.hot=true`가 된다.
+- 기본값: `minSentences=4`, `minEojeols=20`, `spacing.eojeolLengthCV <= 0.38`,
+  `comma.perSentence < 1`, `posProxy.matchedCount >= 10`,
+  `posProxy.classDiversity <= 0.26`.
+- 쉼표가 적다는 사실만으로 hot 처리하지 않는다. spacing과 suffix diversity가 동시에
+  맞아야 한다.
+- `.patina.default.yaml`의 `stylometry.ko_diagnostics.enabled=false`로 이 composite만 끌 수
+  있다. 기존 burstiness/MATTR/lexicon 신호는 그대로 동작한다.
+
+Calibration note (2026-05-22): `npm run benchmark:katfish-ko -- --write --basename katfish-ko-latest` evaluates this zero-dependency composite against a private KatFish checkout plus the 250-row public-web KO human-control set. Current aggregate result: KatFish catch rate improves from 58.9% without KO diagnostics to 74.8% with KO diagnostics (+15.9 pp), while the public-web human-control FP rate stays 42/250 (16.8%, +0 rows). The KatFish raw rows are not committed because the upstream repository does not expose repo-level license metadata; only aggregate reports are tracked.
+
+---
+
 ## 6. Hot Decision Rule
 
-단락 수준 SUSPECT 판정은 단순한 OR 규칙으로 결정한다.
+단락 수준 SUSPECT 판정은 단순한 OR 규칙으로 결정한다. v3.7 이후 lexicon density가
+세 번째 hot 신호로 추가되었다(상세 calibration은 §16).
 
 ```
 paragraph is SUSPECT iff
-  burstiness_band == "low"  OR  MATTR_band == "low"
+  burstiness_band == "low"  OR  MATTR_band == "low"  OR
+  (lexicon_density > threshold AND lexicon_min_hits is satisfied)  OR
+  koDiagnostics.hot == true
 ```
 
 ### 근거
 
-- 두 신호 중 하나만 약해도 사람 글에서는 드물게 동시 발생한다
+- 네 신호 중 하나만 약해도 우선 검토할 만한 edit hotspot이 된다
 - AND 조건은 recall 이 너무 낮아 v1 acceptance criteria(AI 시드 7/10)를 만족하기 어렵다
 - false positive 는 `FalsePositiveControl` 평가(자연 시드 ≤2/10)로 견제한다
+- ko/zh/ja lexicon은 단일 hit를 hot 판정으로 쓰지 않고 audit hint로만 남긴다. 짧은
+  한국어 전문/정책 문단에서 `기반`, `흐름` 같은 보통 명사가 단독으로 터지는 오탐을
+  막기 위해 CJK 기본 `lexicon_min_hits`는 2다.
+- ko diagnostic signal은 내부에서 AND composite를 사용하므로 쉼표/조사 같은 단일 특징만으로는
+  OR-rule에 들어오지 않는다
 
 ### 임계값 재조정
 
 v1 결과에서 false positive 가 한도(자연 시드 2/10)를 넘기면, `.patina.default.yaml`의
 `stylometry.burstiness.bands.low` 또는 `stylometry.ttr.bands.low` 를 보수적으로 낮춰
-재평가한다 (예: 0.30 → 0.27).
+재평가한다 (예: 0.30 → 0.27). lexicon 쪽은 `lexicon.density_threshold`로 조정한다.
+ko composite는 `stylometry.ko_diagnostics.bands`에서 조정한다.
 
 ---
 
@@ -244,7 +294,7 @@ emit groups of length ≥ 2 as sub-flags
 |------|------|
 | 단락 수 ≤ 2 | 4.6단계 전체 skip — meta block 생략 |
 | 전체 문장 수 ≤ 2 | 4.6단계 전체 skip — meta block 생략 |
-| 단락 내 문장 수 < 2 | 해당 단락은 burstiness 계산 불가 → MATTR 만 평가 |
+| 단락 내 문장 수 < 3 | CV 값은 기록하지만 burstiness band/hot 판정은 생략 → MATTR/lexicon 만 평가 |
 | 단락 내 토큰 수 = 0 | 해당 단락 skip |
 | 언어가 `stylometry.languages` 에 없음 | 4.6단계 전체 skip (기본값은 ko/en/zh/ja 포함) |
 
@@ -441,8 +491,9 @@ optional dependency로 설치 실패가 없는 경로, (2) ko/en/zh/ja benchmark
 | ~~n-gram redundancy~~ | ~~bi/trigram 반복도~~ — **dropped, §15 negative finding** | — |
 | ~~AI-lexicon overlap~~ | ~~28-패턴 외 AI 특유 어구 사전 매칭~~ — **shipped v3.7, §16** | v3.7 |
 | ~~zh/ja character fallback~~ | ~~Han/Kana character-token burstiness/TTR~~ — **shipped for 4.6** | current |
+| ~~ko diagnostic composite~~ | ~~spacing/comma/suffix proxy~~ — **shipped as conservative AND signal, §5.1** | current |
 | Perplexity proxy | small LM 또는 cloze prompt 기반 token-level surprise | v4 후보 |
-| Function-word distribution | 기능어 빈도 분포 차이 (Mosteller & Wallace 류) | v4 후보 |
+| Broader function-word distribution | en/zh/ja 포함 기능어 빈도 분포 차이 | v4 후보 |
 | GPTZero / Originality 연동 | 외부 detector API 결과를 hot 신호로 합산 | v4+ 후보 |
 | 형태소 분석 기반 ko 토큰화 | 어절 → 형태소 단위로 정밀화 | v2+ |
 | zh/ja 형태소 분석 통합 | 단어 경계 처리; 위 go/no-go 충족 시에만 재검토 | v4+ 후보 |
@@ -456,6 +507,9 @@ optional dependency로 설치 실패가 없는 경로, (2) ko/en/zh/ja benchmark
 - **Korean morphology coarseness**: 어절 단위 토큰화는 morpheme-level 분석과 다르다.
   같은 명사의 조사 변형(`도구는`, `도구가`, `도구를`)을 서로 다른 token 으로 취급하므로
   MATTR 이 실제보다 약간 높게 나올 수 있다.
+- **Korean diagnostic composite is conservative**: 쉼표 없음 + 균질한 어절 길이 + 낮은 suffix
+  class diversity가 모두 맞아야 하므로 recall보다는 false-positive 억제를 우선한다. 실제
+  KatFish/2025+ corpus 재보정 전에는 넓은 성능 주장으로 쓰지 않는다.
 - **Short text noise**: 단락 ≤2 또는 문장 ≤2 인 텍스트는 통계적으로 신뢰할 수 없어 skip
   한다. 이 경우 4.6단계 전체가 비활성화된다.
 - **Window=50 fallback**: MATTR window 보다 짧은 단락은 simple TTR 로 fallback 하므로
@@ -467,8 +521,8 @@ optional dependency로 설치 실패가 없는 경로, (2) ko/en/zh/ja benchmark
 - **False split**: 약어(`Mr.`, `e.g.`) 와 소수점(`3.14`) 같은 종결부호 오인식을 v1 에서
   감수한다. 시드 평가에서 false positive 비율이 한도를 넘기면 v1.1 에서 보강한다.
 - **Language scope**: 4.6 stylometry는 ko/en/zh/ja를 지원한다. zh/ja는 형태소가 아니라
-  character-token fallback이므로 MATTR 해석은 보수적으로만 사용한다. 4.7 lexicon은
-  curated 사전이 있는 en/ko만 기본 지원한다.
+  character-token fallback이므로 MATTR 해석은 보수적으로만 사용한다. 4.7 lexicon도
+  en/ko/zh/ja 기본 사전을 제공한다.
 
 ---
 
@@ -612,33 +666,37 @@ For paragraph P with tokens T:
              substring for "Multi-word phrases")
   density = matches / len(T) * 1000   # matches per 1000 tokens
 
-  hot iff density > threshold
+  hot iff density > threshold AND min_hot_matches is satisfied
 ```
 
 기본 threshold = `2.0` (1,000 토큰당 2회). `lexicon.density_threshold`로 설정 가능.
+기본 `min_hot_matches`는 영어 1, 한국어/중국어/일본어 2다. CJK 단일 lexicon hit는
+audit hint로 표시하지만 단락을 hot으로 만들지는 않는다.
 
-### 단락 hot 결정 규칙 (3-signal OR)
+### 단락 hot 결정 규칙 (v3.7 당시 3-signal OR)
 
 ```
 paragraph is SUSPECT iff
-  burstiness_band == "low" OR MATTR_band == "low" OR lexicon_density > threshold
+  burstiness_band == "low" OR MATTR_band == "low" OR
+  (lexicon_density > threshold AND min_hot_matches is satisfied)
 ```
 
-§6의 2-signal OR 규칙을 3-signal OR로 확장한다. burstiness/MATTR는 분포적 신호, lexicon은 어휘적 신호 — 다른 축이라 OR 결합 시 둘 다 합산 효과를 낸다.
+v3.7에서는 §6의 2-signal OR 규칙을 3-signal OR로 확장했다. burstiness/MATTR는 분포적 신호, lexicon은 어휘적 신호 — 다른 축이라 OR 결합 시 둘 다 합산 효과를 냈다. 현재 전체 규칙은 §6의 4-signal OR이며, ko diagnostic composite가 네 번째 축이다.
 
 ### 사전 파일
 
-- `lexicon/ai-en.md` — 영어 50개 strict + 58개 phrase = 108 entries
-- `lexicon/ai-ko.md` — 한국어 41개 strict + 49개 phrase = 90 entries
+- `lexicon/ai-en.md` — 영어 43개 strict + 45개 phrase = 88 entries
+- `lexicon/ai-ko.md` — 한국어 43개 strict + 53개 phrase = 96 entries
+- `lexicon/ai-zh.md` — 중국어 60개 phrase = 60 entries
+- `lexicon/ai-ja.md` — 일본어 60개 phrase = 60 entries
 
-탐색은 `Glob lexicon/ai-{lang}.md`로 자동. zh/ja는 4.6 stylometry만 기본 지원하고
-4.7 lexicon은 curated 사전이 없어 기본 미지원(`lexicon.languages` 기본 `[en, ko]`).
-사용자가 `custom/lexicon/ai-{lang}.md` 를 두면 우선 로드.
+탐색은 `Glob lexicon/ai-{lang}.md`로 자동. 기본 `lexicon.languages`는
+`[en, ko, zh, ja]`이다. 사용자가 `custom/lexicon/ai-{lang}.md` 를 두면 우선 로드.
 
 ### 매칭 정책
 
-- **Strict matches** (Markdown `## Strict matches` 섹션): 대소문자 무시 whole-word 매칭. 한국어 어절(예: `자리매김`)은 substring으로 근사 — Korean punctuation/space로 분리되므로 false attach 위험이 낮다
-- **Multi-word phrases** (Markdown `## Multi-word phrases` 섹션): 대소문자 무시 substring 매칭. `~` 가 포함된 한국어 phrase(예: `~의 지평을 넓히다`)는 `~` 을 wildcard 로 취급 (`.{0,40}`)
+- **Strict matches** (Markdown `## Strict matches` 섹션): 대소문자 무시 whole-word 매칭. CJK(ko/zh/ja)는 substring으로 근사 — 한국어 어절과 zh/ja character fallback에서 multi-character entry가 통째 token으로 남지 않기 때문이다
+- **Multi-word phrases** (Markdown `## Multi-word phrases` 섹션): 대소문자 무시 substring 매칭. `~` 가 포함된 phrase(예: `~의 지평을 넓히다`, `不仅~而且`, `~と言えるでしょう`)는 `~` 을 wildcard 로 취급 (`.{0,40}`)
 - 한 paragraph 안에서 같은 entry 가 여러 번 나와도 1로 계산 (entry 단위 카운트)
 
 ### LLM 전달 형식 확장
@@ -694,7 +752,7 @@ Pareto frontier (3-signal OR, threshold sweep):
 
 ### Threshold 선택 근거
 
-`density_threshold = 2.0` 채택. 0.5–5.0 plateau 구간 어디에서도 동일한 catch/FP 가 나오므로 사양 기본값(2.0) 을 사용한다. 운용 의미: "1,000 토큰당 AI lexicon entry 가 2개 초과로 나타나면 단락 의심". 이는 사양 §3 Recommendation 과 일치한다.
+`density_threshold = 2.0` 채택. 0.5–5.0 plateau 구간 어디에서도 동일한 catch/FP 가 나오므로 사양 기본값(2.0) 을 사용한다. 운용 의미: "1,000 토큰당 AI lexicon entry 가 2개 초과로 나타나고, 언어별 최소 hit 수를 만족하면 단락 의심". 이는 사양 §3 Recommendation 과 일치한다. 2026-05 Korean 25-row register pilot 이후 CJK 단일 hit는 hot에서 audit hint로 낮췄다.
 
 ### Calibration drop list
 
@@ -714,7 +772,17 @@ lexicon 은 다음 카탈로그 항목과 **중복되지 않도록** 큐레이
 - `ko-language.md` Pattern 7, 8 (다양한, 활발한, 혁신적인, ~적 접미사 등) — 카탈로그가 이미 다룸
 - `ko-style.md` Pattern 13, 18 (이를 통해, 도모하다, 본 사업 등) — 카탈로그가 이미 다룸
 
-lexicon 의 50+58/41+49 entry 는 위 패턴들이 명시적으로 잡지 않는 영역(modal scaffolding, 추상명사, 의례적 도입/마무리 phrase) 만 추렸다.
+lexicon 의 en 88 / ko 96 / zh 60 / ja 60 entry 는 위 패턴들이 명시적으로 잡지 않는 영역(modal scaffolding, 추상명사, 의례적 도입/마무리 phrase) 만 추렸다.
+
+### zh/ja lexicon 스타터 팩 (v3.11.x)
+
+`lexicon/ai-zh.md`와 `lexicon/ai-ja.md`는 각각 60개 phrase-only entry로 시작한다.
+두 언어는 whitespace token이 안정적이지 않으므로 모든 기본 entry를 `Multi-word phrases`
+섹션에 두고 substring/wildcard 매칭으로 처리한다. 회귀 fixture에는 lexicon-only hot
+샘플과 natural cold counterexample을 각 언어별로 추가했으며, `npm run benchmark` 기준
+zh/ja 모두 false positive 0/4, AI catch 4/4를 유지한다. 더 큰 외부 corpus로 확장할 때는
+각 파일의 counterexample table을 먼저 검토하고, 백과사전·뉴스 register에서 비슷하게
+발화하는 entry는 drop한다.
 
 ### 한국어 lexicon 의 검증 결과 + v3.8.0 재큐레이션
 
@@ -727,6 +795,9 @@ ko/AI 코퍼스 vs NamuWiki human 차별 빈도 마이닝(`.omc/research/v3_8_ko
 - Strict (8개): `평가된다`, `꼽힌다`, `가리킨다`, `사례로`, `다수의`, `알려져`, `일컬어진다`, `평가받다`
 - Phrases (4개): `가운데 하나로`, `자리 잡았다`, `알려져 있다`, `~의 사례로`
 
+**v3.12.x false-positive pruning (96 entries)**:
+`환경`, `기반`, `흐름`, `측면`, `토대`, `가리킨다`는 기술 문서의 구체적 용례(`환경 변수`, `이벤트 기반`, `인증 흐름` 등)에서 짧은 단락을 과열시키는 bare strict entry라 기본 lexicon에서 제외했다. 같은 register 과잉은 더 구체적인 phrase나 retained marker(`자리매김`, `중요한 의미`, `생태계`, `양상` 등)로 잡는다.
+
 **v3.8.0 결과** (동일 코퍼스 재측정, `.omc/research/v3_8_remeasure.py`):
 
 | Source | n | v3.7.0 hot | v3.8.0 hot | Δ | lex fires |