patina-cli 3.11.0 → 4.0.0

package/docs/PATTERNS-ZH.md

@@ -2,8 +2,8 @@
 
 This page expands the Chinese pattern packs into a browsable reference. It is generated from `patterns/zh-*.md`, so the numbers, names, watch words, fire conditions, and examples mirror the source pattern files.
 
-- Rewrite-capable patterns: 32
-- Score/audit-only viral-hook patterns: 8
+- Rewrite-capable patterns: 33
+- Score/audit-only viral-hook patterns: 9
 - Main selector: [PATTERNS.md](PATTERNS.md)
 
 ## Pattern Index
@@ -23,6 +23,7 @@ This page expands the Chinese pattern packs into a browsable reference. It is ge
 | 11 | rewrite | 同义词循环替换 | [zh-language.md](../patterns/zh-language.md) |
 | 12 | rewrite | 冗长的介词结构 | [zh-language.md](../patterns/zh-language.md) |
 | 32 | rewrite | “更”比较副词滥用 | [zh-language.md](../patterns/zh-language.md) |
+| 33 | rewrite | 定义式隐喻等式（"X 是 Z 的[抽象名词]"） | [zh-language.md](../patterns/zh-language.md) |
 | 13 | rewrite | 过度使用连接词/过渡词 | [zh-style.md](../patterns/zh-style.md) |
 | 14 | rewrite | 加粗滥用 | [zh-style.md](../patterns/zh-style.md) |
 | 15 | rewrite | 内联标题列表 | [zh-style.md](../patterns/zh-style.md) |
@@ -50,6 +51,7 @@ This page expands the Chinese pattern packs into a browsable reference. It is ge
 | VH-6 | score/audit only | 伪统计引用 | [zh-viral-hook.md](../patterns/zh-viral-hook.md) |
 | VH-7 | score/audit only | 头衔堆叠式权威 | [zh-viral-hook.md](../patterns/zh-viral-hook.md) |
 | VH-8 | score/audit only | 未来自我 / 拟亲密二人称承诺 | [zh-viral-hook.md](../patterns/zh-viral-hook.md) |
+| VH-9 | score/audit only | 格言式收尾 / 独立判断短句（伪深刻"金句"） | [zh-viral-hook.md](../patterns/zh-viral-hook.md) |
 
 ## 内容模式
 
@@ -252,7 +254,7 @@ Example after:
 - Source: [zh-language.md](../patterns/zh-language.md)
 - Type: rewrite-capable pattern
 - Watch words: 更具体、更高效、更全面、更深入、更清晰、更积极、更系统、更有效、更稳健、更广泛、更优质、更可持续、更具战略性
-- Fire condition: 同一文档中出现2个以上“更 + 形容词/副词/抽象名词化短语”，且没有明确比较对象、基准或指标。
+- Fire condition: 同一文档中出现2个以上“更 + 形容词/副词/抽象名词化短语”，且没有明确比较对象、基准或指标。若同一段落中只有1处“更X”，但同时出现“系统推进”“全面提升”“战略框架”“协同机制”等正式AI标记，也可以触发。
 - Example files: [failure](../examples/zh-32-failure-01.md) · [success](../examples/zh-32-success-01.md)
 
 Example before:
@@ -263,6 +265,36 @@ Example after:
 
 > 下一阶段先做三件事：把里程碑拆到每周，给每个任务指定负责人，重新核对预算表。产品、运营和法务每周二开30分钟会，专门处理跨部门卡住的事项。
 
+### 33. 定义式隐喻等式（"X 是 Z 的[抽象名词]"）
+
+- Source: [zh-language.md](../patterns/zh-language.md)
+- Type: rewrite-capable pattern
+
+**注意词汇（抽象名词标记）：** 签名、印记、注脚、形状、轮廓、语言、货币、架构、骨架、引擎、心跳、脉搏、命脉、底色、基石、基因/DNA、底层逻辑、本质、缩影
+
+**问题：** AI 中文喜欢用一个系动词判断句把两个东西强行划等号——"X 是 Z 的[抽象名词]"——借宏大的隐喻制造深刻感。"对称是信任的架构""尴尬是你主动选择的那条梯度留下的可见印记"之类的句子听上去很有哲理，实则只是把一个抽象概念套在另一个上面，没有给出任何可验证的支撑。它常常单独成句，充当段落的"金句"。
+
+**与 #8（系动词回避）的区分（务必不要混淆）：** #8 标记的是**回避"是"**——把"X 是 Y"写成"X 充当 Y""X 起到 Y 的作用""X 相当于 Y"，修改方向是**改回"是"**。#33 恰好相反：它标记的是已经用了"是"、却把"是"膨胀成"X 是 Z 的[抽象名词]"这种**夸张的隐喻等式**，修改方向是**拆掉这个等式、换成具体陈述**。一个是缺"是"，一个是"是"被滥用造句，两者不是同一回事。
+
+**触发条件：** 同一文档/段落中出现 2 处以上“X 是 Z 的[抽象名词]”形式的判断句（抽象名词取自上方标记词或同类），这些等式依赖宏大隐喻而不是具体证据成立。单次出现只记作 audit 提示，不直接进入 rewrite。
+
+- Example files: [failure](../examples/zh-33-failure-01.md) · [success](../examples/zh-33-success-01.md)
+**排除条件：**
+- 字面或技术性定义（"水是万能溶剂""线粒体是细胞的能量工厂""哈希表是一种以键值映射为基础的数据结构"）
+- 约定俗成的固定说法或事实性等同（"北京是中国的首都""黄河是中华民族的母亲河"作为既有惯用语）
+- 后文用具体事实、数据或机制真正撑起这个等式的等同关系（不是空泛比喻，而是有论据的论断）
+
+**Semantic Risk:** MEDIUM
+**Preservation Note:** 拆解隐喻等式时，先判断作者到底想说什么真实主张：是"Z 依赖 X"、"X 能预测 Z"，还是"X 是 Z 的前提条件"？保留这个真实关系和它的方向、强度，只去掉"是…的[抽象名词]"这层装饰；若原文除了比喻之外没有任何可保留的信息，就改成一个具体的事实、动作或观察。
+
+**修改前：**
+> 对称是信任的架构。在产品设计里，一致性是用户体验的命脉，而留白则是高级感的语言。
+
+**修改后：**
+> 界面左右对齐、控件间距统一时，用户更容易预判下一步操作，投诉里"找不到按钮"的比例下降了。我们在两版改版里都验证过这一点。
+
+---
+
 ## 风格模式
 
 ### 13. 过度使用连接词/过渡词
@@ -528,6 +560,7 @@ Example after:
 - Type: rewrite-capable pattern
 - Watch words: 这是一个～的事实、可以被认为是、作为一种～的方式、被～所～、对于～来说、基于～的原因、关于～的问题、通过～来实现
 - Fire condition: 同一段落中出现2个以上翻译腔表达。单次使用可容忍——正常中文偶尔也会出现。
+- Context guard: issue #352 — when translationese appears around em dash/colon/slash/comma splice punctuation, do not replace only the punctuation; rebuild the clause relation in natural Chinese, e.g. `无 TUI 设置` → `不用 TUI 就完成全自动安装...`.
 - Example files: [failure](../examples/zh-26-failure-01.md) · [success](../examples/zh-26-success-01.md)
 
 Example before:
@@ -704,3 +737,34 @@ Detection example:
 > 改写前：朋友，先收藏这篇。一年后的你一定会感谢现在的自己。
 >
 > 改写后：如果你下个月要做计划，可以保存这份清单。
+
+### Viral 9. 格言式收尾 / 独立判断短句（伪深刻"金句"）
+
+- Source: [zh-viral-hook.md](../patterns/zh-viral-hook.md)
+
+**关注词汇：**（结构性模式——按形式判断，不看具体词汇）
+
+**问题：** 一句很短（约 10 字以内）、语法完整的判断句被单独放在一行或一段，借这种"留白 + 收束"的排版制造庄重感和深刻感——典型的伪深刻"金句""mic drop"。一篇里出现 2 句以上，或每段都用这样一句收尾时，就成了 AI 网红长文的强信号。它靠的是**形式**（独立、短、斩钉截铁），而不是某个特定词汇。
+
+**触发条件：** 出现独立成行/成段、语法完整、约 10 字以内的判断式短句，且其作用是修辞性收束（金句、点题），而非传递新信息。一篇出现 1 句即可记为低信号，2 句及以上信号显著。
+
+**严重度标尺：**
+- Low：全篇仅 1 句独立格言短句。
+- Medium：出现 2 句独立格言短句。
+- High：出现 3 句及以上，或几乎每个段落都以一句这样的短句收尾。
+
+**排除条件：**
+- 诗、歌词、韵文
+- 本就简短的备忘、通知、提示、对问题的一两句回答
+- 有意为之且后文有具体支撑的格言（不是空泛点题，而是引出论据）
+- 对话、引语
+- 标题、小标题
+
+**Semantic Risk:** LOW —— 仅评分，重写时不动。
+**Preservation Note:** 该模式默认仅评分不重写；若用户明确要求降调，保留这句话真正想表达的判断或重点，把它并回相邻段落或补上具体支撑，而不是简单删掉，以免抹掉作者的核心观点。
+**改写前 / 改写后示例（手动降信号）：**
+> 改写前：增长是结果，不是目标。
+> 真正重要的从来不是数字。
+> 对称会变成陷阱。
+>
+> 改写后：我们今年没有把月活当成首要指标，而是盯着留存率和复购，因为一味追规模上次让我们烧掉了两个季度的预算。

package/docs/PATTERNS.md

@@ -1,13 +1,13 @@
 # Pattern Catalog
 
-Patina ships 160 pattern entries across four languages. The language-specific references below expand each pack with pattern numbers, names, watch words, fire conditions, source links, and examples.
+Patina ships 168 pattern entries across four languages. The language-specific references below expand each pack with pattern numbers, names, watch words, fire conditions, source links, and examples.
 
 | Language | Reference | Rewrite-capable patterns | Score/audit-only viral-hook patterns |
 |----------|-----------|--------------------------|--------------------------------------|
-| Korean | [PATTERNS-KO.md](PATTERNS-KO.md) | 32 | 8 |
-| English | [PATTERNS-EN.md](PATTERNS-EN.md) | 32 | 8 |
-| Chinese | [PATTERNS-ZH.md](PATTERNS-ZH.md) | 32 | 8 |
-| Japanese | [PATTERNS-JA.md](PATTERNS-JA.md) | 32 | 8 |
+| Korean | [PATTERNS-KO.md](PATTERNS-KO.md) | 33 | 9 |
+| English | [PATTERNS-EN.md](PATTERNS-EN.md) | 33 | 9 |
+| Chinese | [PATTERNS-ZH.md](PATTERNS-ZH.md) | 33 | 9 |
+| Japanese | [PATTERNS-JA.md](PATTERNS-JA.md) | 33 | 9 |
 
 ## Notes
 

package/docs/RESEARCH-DOCS-PLATFORM.md

@@ -0,0 +1,54 @@
+# Docs Platform RFC
+
+Status: investigation only; no docs-site implementation in this issue.
+Sources checked: 2026-05-20 UTC.
+
+Patina currently ships flat Markdown plus generated API/pattern reference pages. That remains the lowest-maintenance option for the next release. A docs site becomes worthwhile when search, localized navigation, or versioned release docs become more valuable than keeping every document directly browsable in GitHub.
+
+## Options compared
+
+| Option | i18n | Search | Versioning | Deployment effort | Notes |
+|---|---|---|---|---|---|
+| Docusaurus | Strong built-in locale workflow and translated docs routes | Built-in local search options and hosted search integrations | Built-in docs versioning | Medium: React app, sidebars, build pipeline, deploy to Pages/Vercel/Netlify | Best if Patina wants versioned docs and a conventional OSS docs portal. |
+| Astro Starlight | Strong i18n routing for content collections | Built-in Pagefind search in the Starlight docs stack | No first-class versioned-docs workflow comparable to Docusaurus; can model versions as content sections | Medium-low: Astro app plus content migration | Best fit if Patina wants a lightweight multilingual docs site with modern content authoring and fewer React conventions. |
+| MkDocs + Material | Mature Markdown docs, search, navigation, and language customization; i18n usually needs plugin/workflow choices | Built-in client search through MkDocs/Material | Usually handled with `mike` or deploy aliases | Medium: Python toolchain plus theme/plugin choices | Good for Python-heavy teams; less natural for this Node-first repo. |
+| GitHub Pages with current Markdown | No routed i18n beyond separate files | GitHub search only; no site search UX | Tags/branches only, no docs-version UX | Low: already works | Best current default. Keeps maintenance near zero but does not solve navigation/search. |
+
+## Source links
+
+- Docusaurus docs: https://docusaurus.io/docs
+- Docusaurus i18n: https://docusaurus.io/docs/i18n/introduction
+- Docusaurus versioning: https://docusaurus.io/docs/versioning
+- Astro Starlight docs: https://starlight.astro.build/
+- Starlight i18n: https://starlight.astro.build/guides/i18n/
+- Starlight search: https://starlight.astro.build/guides/search/
+- MkDocs: https://www.mkdocs.org/
+- Material for MkDocs: https://squidfunk.github.io/mkdocs-material/
+- Material search: https://squidfunk.github.io/mkdocs-material/setup/setting-up-site-search/
+- `mike` MkDocs versioning: https://github.com/jimporter/mike
+- GitHub Pages: https://docs.github.com/pages
+
+## Recommendation
+
+Do **not** migrate immediately. Keep GitHub-rendered Markdown until one of these triggers is true:
+
+1. search becomes a recurring support problem;
+2. localized docs need language-specific navigation rather than separate `README_*` files;
+3. a release process requires versioned docs for multiple supported CLI versions.
+
+When a site is justified, start with **Astro Starlight** unless versioned docs are the primary requirement. Starlight is the best fit for Patina's current content shape: Markdown-first, multilingual, and lightweight. Choose **Docusaurus** instead if versioning becomes mandatory before the migration starts.
+
+## Effort estimate
+
+| Phase | Estimate | Work |
+|---|---:|---|
+| RFC acceptance + URL plan | 0.5 day | Pick `/docs` vs custom domain, define canonical source of truth, decide whether README remains the landing page. |
+| Minimal Starlight proof of concept | 1-2 days | Add app scaffold, import 8-12 key docs, configure i18n nav/search, deploy preview. |
+| Full migration | 3-5 days | Move docs into site collections, preserve links/anchors, add redirects, wire CI build, update README links. |
+| Versioned docs with Docusaurus | 4-7 days | Same as above plus sidebar/version policy and release tagging workflow. |
+
+## Non-goals
+
+- No docs-site dependency is added by this RFC.
+- No README or existing Markdown page is moved.
+- No custom domain, Pages workflow, or search index is created here.

package/docs/ROADMAP.md

@@ -13,12 +13,14 @@ This roadmap focuses on two things:
 - Public scope: Korean, English, Chinese, Japanese AI-writing pattern rewriting
 - Current benchmark layer:
   - deterministic stylometry/lexicon benchmark: `npm run benchmark`
+  - adversarial MPS fixture gate: `npm run quality:adversarial-mps`
+  - 2026 rebaseline status: [`docs/research/2026-rebaseline.md`](research/2026-rebaseline.md)
 - Current public calibration claim:
-  - Korean editing-hotspot recall: 91% [84.0-95.4%], n=100
-  - English HC3 editing-hotspot recall: 76% [66.7-83.3%], n=100
-  - human false positives: 13-25% point-estimate range across registers
-- Current distribution gap:
-  - package name `patina-cli` is not published on npm as of 2026-05-20
+  - 2026-05-22 modern-model catch: 67.3% [63.5-71.0%], n=600 across KO+EN × GPT/Claude/Gemini
+  - human-control false positives: 16.0% [11.6-21.7%], n=200 across KO+EN
+  - per-cell results: `docs/benchmarks/rebaseline-latest.md`
+- Current distribution:
+  - npm package `patina-cli` is the public distribution channel; repo metadata currently targets `4.0.0` and should be verified with `npm run release:check` before publishing.
 
 ## 0. Positioning principles
 
@@ -58,14 +60,16 @@ Avoid overclaiming:
 Goal: make claims easier to verify and harder to dismiss.
 
 - Publish a short benchmark report generated from `tests/quality/results.json`.
-- Add ROC-AUC / PR-AUC and threshold sweep to the deterministic benchmark.
+- Keep ROC-AUC / PR-AUC and threshold sweep diagnostics current in the deterministic benchmark report.
 - Split reports by language, class, and register.
 - Add a visible warning that scores measure AI-likeness, not authorship.
 - Link [`docs/research/ai-human-metrics.md`](research/ai-human-metrics.md) from README.
+- Keep the adversarial MPS report current so high meaning preservation cannot hide unchanged AI-like style.
 
 Acceptance criteria:
 
 - `npm run benchmark` still passes.
+- `npm run quality:adversarial-mps` still passes.
 - Benchmark output includes current binary metrics plus ranked/threshold metrics.
 - README claims are traceable to a specific benchmark report or spec section.
 
@@ -250,66 +254,42 @@ Do not lead with “bypass AI detectors.” Lead with:
 
 ## 5. Immediate next actions
 
-Last triaged: 2026-05-20, after Wave 1 docs/governance quick wins.
+Last triaged: 2026-06-05, after closing the public launch tracker and moving launch posting notes out of GitHub issues.
 
 Current GitHub issue inventory:
 
-- 42 open issues; 100 closed issues; 142 tracked issues total.
-- Open priority split: 2 high, 26 medium, 14 low, and 0 without priority labels.
-- The only open `priority: high` issues are the blocked release-wave items [#203](https://github.com/devswha/patina/issues/203) and [#204](https://github.com/devswha/patina/issues/204).
-
-Already done or mostly done:
-
-- README links to the roadmap, research notes, benchmark report, detector comparison, demo, launch copy, and community docs.
-- Issue templates, PR template, `SECURITY.md`, `SUPPORT.md`, `CODE_OF_CONDUCT.md`, and `CONTRIBUTING.md` exist.
-- Benchmark report generation exists via `npm run benchmark:report`.
-- Logo, icon, and social preview SVGs live in `assets/brand/` and `assets/social/`.
-- A copy-paste terminal demo lives in [`docs/DEMO.md`](DEMO.md).
-- Launch copy drafts live in [`docs/social/patina-launch-copy.md`](social/patina-launch-copy.md).
-- Main branch protection is enabled. Issue: [#246](https://github.com/devswha/patina/issues/246).
-- MAX warns when every MAX candidate fails `MPS >= 70` and the highest-MPS candidate is selected as fallback. Issue: [#139](https://github.com/devswha/patina/issues/139).
-- Deterministic `callLLM` seams exist across `ouroboros`, `max-mode`, and scoring test paths. Issue: [#130](https://github.com/devswha/patina/issues/130).
-- Localized READMEs have been dogfooded through patina to reduce AI tells. Issue: [#242](https://github.com/devswha/patina/issues/242).
-- zh/ja before-after examples have been backfilled to near parity with ko/en pattern coverage. Issue: [#146](https://github.com/devswha/patina/issues/146).
-- External promotional / solicitation issue policy is documented. Issue: [#245](https://github.com/devswha/patina/issues/245).
-- README hero block is centered. Issue: [#241](https://github.com/devswha/patina/issues/241).
-- Korean companion docs exist for contributing, FAQ, authentication, and examples. Issue: [#202](https://github.com/devswha/patina/issues/202).
-- Priority labels were added for the last unlabeled queue items. Issues: [#99](https://github.com/devswha/patina/issues/99), [#104](https://github.com/devswha/patina/issues/104).
-- Test/backend hardening wave is complete: prompt snapshots, backend contract unification, SIGINT cancellation, structured logging, and progress indicators. Issues: [#169](https://github.com/devswha/patina/issues/169), [#131](https://github.com/devswha/patina/issues/131), [#133](https://github.com/devswha/patina/issues/133), [#132](https://github.com/devswha/patina/issues/132), [#180](https://github.com/devswha/patina/issues/180).
-- Scoring/stylometry quality wave is complete: zh/ja char n-grams, deterministic shadow scoring, manifest v2 observability, response cache, and voice anchors. Issues: [#151](https://github.com/devswha/patina/issues/151), [#136](https://github.com/devswha/patina/issues/136), [#134](https://github.com/devswha/patina/issues/134), [#135](https://github.com/devswha/patina/issues/135), [#137](https://github.com/devswha/patina/issues/137).
-- Pattern/profile wave is complete: zh/ja risk notes, comparison-adverb backport, zh/ja profile overrides, overlap audit, developer-prose profiles, and viral-hook expansion to 8 score-only patterns per language. Issues: [#147](https://github.com/devswha/patina/issues/147), [#148](https://github.com/devswha/patina/issues/148), [#149](https://github.com/devswha/patina/issues/149), [#152](https://github.com/devswha/patina/issues/152), [#153](https://github.com/devswha/patina/issues/153), [#154](https://github.com/devswha/patina/issues/154).
-
-Next executable wave order (snapshot: 2026-05-20 after PR #273):
-
-### Blocked release wave — keep visible, do not start yet
-
-These are the only open high-priority issues. They stay blocked until package ownership, release credentials, versioning, and support policy are settled.
-
-1. Settle npm support policy, then publish `patina-cli` / `patina-humanizer` with release workflow. Issue: [#203](https://github.com/devswha/patina/issues/203).
-2. Ship `devswha/patina-action@v1` for PR-comment scoring after the release/versioning path is stable. Issue: [#204](https://github.com/devswha/patina/issues/204).
-
-### Wave 1 — medium, executable after release policy is unblocked
-
-1. Add zh/ja AI-lexicon files for stylometry overlap detection. Issue: [#104](https://github.com/devswha/patina/issues/104).
-2. Let CLI backends participate in standalone MAX. Issue: [#141](https://github.com/devswha/patina/issues/141).
-3. Reduce MAX prompt weight by opting into a minimal default. Issue: [#143](https://github.com/devswha/patina/issues/143).
-4. Add MAX pane-liveness watchdogs for dead tmux panes. Issue: [#144](https://github.com/devswha/patina/issues/144).
-5. Make `patina auth login <backend>` launch the real login flow. Issue: [#186](https://github.com/devswha/patina/issues/186).
-6. Add JSDoc public exports and publish generated API reference. Issue: [#191](https://github.com/devswha/patina/issues/191).
-7. Tie this roadmap to a GitHub Project board or Milestones. Issue: [#195](https://github.com/devswha/patina/issues/195).
-
-### Wave 2 — research calibration, parked unless explicitly scheduled
-
-These are medium-priority but research-heavy. Keep them out of the critical path until release blockers and the core quality loop have stable owner time.
-
-1. Re-baseline AI catch rate against 2025+ models. Issue: [#155](https://github.com/devswha/patina/issues/155).
-2. Run the lexicon freshness audit with per-entry corpus provenance. Issue: [#160](https://github.com/devswha/patina/issues/160).
-3. Establish quarterly pattern-freshness review with corpus refresh and emerging-pattern triage. Issue: [#165](https://github.com/devswha/patina/issues/165).
-
-### Wave 3 — low-priority parked ecosystem and research
-
-Do not start these until npm publication, action v1, and medium MAX/auth/docs work have a stable release path:
-
-- False-positive and benchmark calibration: [#99](https://github.com/devswha/patina/issues/99), [#156](https://github.com/devswha/patina/issues/156), [#157](https://github.com/devswha/patina/issues/157), [#158](https://github.com/devswha/patina/issues/158), [#159](https://github.com/devswha/patina/issues/159), [#163](https://github.com/devswha/patina/issues/163).
-- Documentation site exploration: [#199](https://github.com/devswha/patina/issues/199).
-- Editor/platform integrations and distribution experiments: [#206](https://github.com/devswha/patina/issues/206), [#207](https://github.com/devswha/patina/issues/207), [#208](https://github.com/devswha/patina/issues/208), [#209](https://github.com/devswha/patina/issues/209), [#210](https://github.com/devswha/patina/issues/210), [#211](https://github.com/devswha/patina/issues/211), [#212](https://github.com/devswha/patina/issues/212).
+- 9 open issues.
+- Open PRs: 0.
+- Open priority split: 0 high, 1 medium, 8 low, and 0 without priority labels.
+- No current high-priority issue.
+
+Campaign state:
+
+- Merged campaign PRs: #281, #287, #288, #289, #290, #292, #293.
+- Concurrent cleanup merged during the final gate: #294 closed #291.
+- Final review blocker cleanup: #295.
+- Launch Wave 1 badge work: #297 closed #282; companion patina-action#1 added `badge-json` / `badge-branch`.
+- Launch Wave 2 support work: #299 closed #285; the experimental share-card generator from #283 has since been removed from the CLI surface.
+- Launch Wave 3 static playground work closes #208 and targets <https://patina.vibetip.help/> for the try-it-now URL.
+- Launch execution prep: Korean-first channel drafts live in `docs/social/patina-launch-korean-first.md` and score 0.0%; `docs/social/patina-launch-copy.md` scores 6.3% after the KO diagnostic scoring update. Launch posting/deferral notes are maintainer-owned operational bookkeeping and should be tracked outside public GitHub issues.
+- Rebaseline claim pass: `npm run benchmark:rebaseline:report` refreshes `docs/benchmarks/rebaseline-latest.{md,json}` from the #155 claim-ready sanitized manifest (800 rows, no raw text).
+- KO/2025+ corpus prep: `docs/research/ko-2025-corpus-sources.md` records usable Korean sources, `artifacts/rebaseline-2025/intake.local.example.jsonl` provides the 25-row pilot skeleton, `artifacts/rebaseline-2025/sources.ko-public.jsonl` inventories public Korean web sources, `artifacts/rebaseline-2025/human-controls.public.jsonl` tracks 250 scored hash-only web human-control candidates at n=50 for each tracked register, `npm run benchmark:rebaseline:web` collects raw text into ignored private rows, and `npm run benchmark:rebaseline:score` refreshes deterministic outcome fields without copying raw text.
+- KO register pilot: `npm run benchmark:register-pilot -- --write --basename register-stratified-latest` refreshes false positives by register without committing raw text; the expanded current pilot shows 42/250 predicted-hot human-control rows, split by register for threshold work.
+- KO KatFish calibration: `npm run benchmark:katfish-ko -- --write --basename katfish-ko-latest` reports aggregate-only private KatFish metrics; current KO diagnostics improve catch rate from 58.9% to 74.8% versus Patina without KO diagnostics while public-web human-control FP stays 42/250.
+- Launch feedback prep: the false-positive issue form now captures text origin, redistribution, fired paragraph, score output, and expected behavior.
+- Growth nudge prep: the one-time CLI star reminder from #305 has since been removed to keep stderr operational-only.
+- README demo prep: #306 adds the first-screen terminal GIF, try-it-now
+  playground link, translated README references, and re-recording notes; #308
+  localizes README hero GIF references so English uses `patina-demo-en.gif`,
+  Korean uses `patina-demo-ko.gif`, and ZH/JA intentionally fall back to EN.
+- Closed or verified during the campaign: #99, #104, #155, #156, #157, #160, #303, #165, #186, #191, #199, #209, #210, #286, #304, #305, #306, #308.
+- Kept open with explicit blocker comments or pending external action: #158, #159, #206, #207, #211, #212, #284, #307, #324.
+- Legacy bot/harness notes were removed from the public repo; restart autonomous bot work only from a fresh, tracked design if it becomes necessary.
+
+Next recommended order:
+
+1. Prepare #307 awesome-list discovery submissions only as candidate copy/checklists; maintainer-owned external submissions should stay outside automated repo changes.
+2. Re-implement #324 only when live credentialed quality checks are worth the larger local-runner investment.
+3. Treat low-priority research/ecosystem items (#158, #159, #206, #207, #211, #212, #284) as parked until evaluator budget, reviewer pool, redistributable corpus, external repo, hosting, or governance prerequisites exist; keep new campaign PRs short-lived.
+
+Detailed wave grouping lives in `docs/ISSUE-WAVES.md` so launch/growth, completed profile work, evaluation-gated research, and parked ecosystem work can move independently without re-triage.

package/docs/TRANSLATIONESE-KO.md

@@ -0,0 +1,51 @@
+# 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.
+
+## 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. The SKILL / callers act on it (e.g. the
+  rewrite loop), and the audit surface can display it.
+- **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 기둥**을 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 에이전트**에 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
+
+```js
+analyzeText(text, { lang: 'ko' }).translationese
+// → { count, density, sentences, byRule:[{id,label,strong,count,example}],
+//     hits:[...], hot, thresholds:{count,density} }
+```
+
+## 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.
+- Not wired into `hot` yet — promote only after validating no FP regression on
+  the benchmark / a diverse human ko control set.

package/docs/audits/2026-05-deep-research.md

@@ -1,5 +1,7 @@
 # patina 저장소 심층 분석 보고서
 
+> **Archive note:** 이 보고서는 MAX mode, hosted backend, `--card`, inline `--api-key`, `--save-run`, response cache 제거 전 스냅샷이다. 아래 CLI surface 관찰은 당시 기준이며 현재 도움말/문서의 source of truth가 아니다.
+
 ## 경영진 요약
 
 `devswha/patina`는 방향이 분명한 저장소다. 패턴 기반으로 AI 글쓰기 흔적을 탐지하고, 감사 가능한 방식으로 다시 쓰기를 수행하며, 한국어·영어·중국어·일본어에 대해 총 126개 패턴을 운영한다. 공개 문서와 코드 기준으로 보면, 이 도구는 standalone Node CLI이자 entity["company","Anthropic","ai company"] Claude Code·entity["company","OpenAI","ai company"] Codex CLI 계열 워크플로에 연결되는 스킬형 도구로 설계되어 있고, rewrite·audit·score·diff·ouroboros·MAX 실행 경로를 이미 갖추고 있다. 또한 backend와 provider를 분리했고, 패턴 팩·프로필·voice·scoring 문서를 프롬프트에 명시적으로 주입하는 구조라 추적 가능성도 좋다. citeturn18view0turn18view1turn26view6turn15view0turn14view2turn23view13
@@ -51,7 +53,7 @@ flowchart LR
 | combined score | `combinedScore`는 `ai-likeness`와 `100-fidelity`를 가중합으로 계산하며 profile별 weight도 config에 있다. | 계산해 놓고도 현재 제어 로직에서 핵심 decision variable로 쓰이지 않는다. design intent와 runtime control이 어긋나 있다. citeturn13view4turn23view10turn27view5 |
 | MAX orchestration | MAX는 각 후보를 생성한 뒤 `scoreText`와 `scoreMPS`를 모두 `models[0]`으로 재채점하고, MPS 원문은 `prompt.split('## Input Text')[1]`로 뽑는다. | 평가 편향과 파싱 취약성이 동시에 존재한다. 생성 모델이 여러 개인데 평가기가 사실상 하나이며, source text 전달도 brittle하다. citeturn28view1turn28view3 |
 | Ouroboros | Ouroboros는 target score, plateau, fidelity floor, MPS floor, rollback을 갖고 있지만, 중단 판단은 `currentScore`, `delta`, floor 위주다. `combined`는 로그에만 남는다. | iterative refinement를 하면서도 실제 결정은 단순 AI-likeness 중심이라, 의미 보존과 자연스러움 사이 균형을 충분히 활용하지 못한다. citeturn14view8turn27view0turn27view3turn27view5 |
-| multilingual | 패턴 카탈로그는 4개 언어를 지원하지만, README의 stylometric/AI-lexicon 설명은 EN 약 108개, KO 약 102개만 명시한다. | zh/ja에 대해 pattern 지원은 있어도 stylometry/lexicon calibration은 약할 가능성이 높다. 특히 paraphrase·translation·summarization처럼 semantic-invariant 과제는 탐지가 더 어렵다는 외부 연구와도 맞물린다. citeturn18view0turn18view4turn33view1 |
+| multilingual | 패턴 카탈로그는 4개 언어를 지원하지만, README의 stylometric/AI-lexicon 설명은 EN 88개, KO 102개, ZH/JA 각 60개를 명시한다. | zh/ja에 대해 pattern 지원은 있어도 stylometry/lexicon calibration은 약할 가능성이 높다. 특히 paraphrase·translation·summarization처럼 semantic-invariant 과제는 탐지가 더 어렵다는 외부 연구와도 맞물린다. citeturn18view0turn18view4turn33view1 |
 | testing/CI | `npm test` 엔트리포인트는 현재 `node --test tests/e2e/*.test.js`만 가리킨다. | e2e 중심 smoke는 가능해도, 점수식·파서·패턴 스키마·Unicode edge case를 잡는 unit/property/quality benchmark 층이 약하다. citeturn44view0turn44view1 |
 | security | 공개 help는 `--api-key <key>`와 여러 provider 환경변수를 직접 안내한다. | 명령행 인자는 shell history나 process list에 노출될 수 있어 비밀정보 전달 수단으로 바람직하지 않다. GitHub 문서와 CWE도 command line 전달을 피하라고 권고한다. citeturn45view0turn45view1turn41search1turn41search3 |
 | reproducibility/UX | 현재 help에는 `--json`, `--save-run`, prompt hash, manifest, config hash 같은 실험 기록 옵션이 보이지 않는다. | README가 score variance를 인정하고 있어도 사용자가 어떤 model/provider/pattern set으로 결과가 나왔는지 재현하기 어렵다. 또한 MAX의 model naming과 provider naming 체계가 분리돼 초심자에게 혼동을 줄 수 있다. citeturn18view3turn23view13turn45view0turn45view1 |

package/docs/benchmarks/README.md

@@ -0,0 +1,51 @@
+# Benchmark reports
+
+This directory stores checked-in benchmark summaries. They are useful for
+regression review, release notes, and public claims only when the matching gate
+says the evidence is ready.
+
+## Files
+
+| File | Source command | Use |
+|---|---|---|
+| `latest.md` / `latest.json` | `npm run benchmark:report` | Deterministic suspect-zone fixture benchmark for KO / EN / ZH / JA, including `signal_score` ROC-AUC / PR-AUC diagnostics. |
+| `detector-comparison.md` / `.json` | `npm run benchmark:compare` | Manual/offline comparison protocol for third-party detectors. |
+| `rebaseline-latest.md` / `.json` | `npm run benchmark:rebaseline:report` | #155 claim-ready 2026 modern-model rebaseline summary (800 hash-only rows; KO+EN × GPT/Claude/Gemini plus human controls). |
+| `katfish-ko-latest.md` / `.json` | `npm run benchmark:katfish-ko -- --write --basename katfish-ko-latest` | Aggregate-only private KatFish calibration for the Korean diagnostic layer; raw rows stay ignored/private. |
+
+## Refresh
+
+```bash
+npm run benchmark:report
+npm run benchmark:compare
+npm run benchmark:rebaseline:report
+npm run benchmark:katfish-ko -- --write --basename katfish-ko-latest
+```
+
+Use `npm run benchmark` for the fast fixture classifier smoke check. Use
+`npm run quality:live` only when you want the opt-in rewrite-quality scaffold;
+by default it does not call a model.
+
+## Public-claim rule
+
+Do not copy numbers into README, launch copy, or social posts unless the report
+itself contains the required evidence. The rebaseline report must stay
+`BLOCKED` until it has scored outcome rows, n≥100 per claim cell, at least two
+languages, at least three generator families, and confidence intervals.
+
+The `latest` report's ranking diagnostics are regression evidence for the
+checked-in fixtures only. They help compare thresholds and signal changes, but
+they are not a general claim that patina detects authorship or current model
+families.
+
+## False-positive loop
+
+If a person-written paragraph is flagged too aggressively, collect it through
+the false-positive form:
+
+<https://github.com/devswha/patina/issues/new?template=false_positive.yml>
+
+A useful report includes the exact paragraph that fired, language/register,
+score output, and whether the sample can become a public fixture. Private or
+vendor-copied text should stay out of the repository; use metadata and hashes
+instead.

package/docs/benchmarks/detector-comparison.json

@@ -1,9 +1,9 @@
 {
   "reportVersion": 1,
-  "generatedAt": "2026-05-20T03:54:28.830Z",
-  "fixtureCount": 34,
-  "benchmarkGeneratedAt": "2026-05-20T03:54:28.820Z",
-  "note": "Offline comparison harness. Built-in Patina row uses deterministic suspect-zone analyzer; third-party rows are manual opt-in only.",
+  "generatedAt": "2026-05-21T15:02:19.159Z",
+  "fixtureCount": 39,
+  "benchmarkGeneratedAt": "2026-05-21T15:02:19.140Z",
+  "note": "Offline comparison protocol. Built-in Patina row uses deterministic suspect-zone analyzer; third-party rows are manual opt-in only.",
   "manualInput": null,
   "detectors": [
     {
@@ -11,17 +11,17 @@
       "name": "Patina deterministic suspect-zone analyzer",
       "kind": "in-tree",
       "mode": "offline",
-      "threshold": "burstiness low OR MATTR low OR lexicon density > threshold"
+      "threshold": "burstiness low OR MATTR low OR lexicon density > threshold with min hits OR koDiagnostics hot"
     }
   ],
   "summaries": {
     "patina-deterministic": {
-      "tp": 18,
+      "tp": 21,
       "fp": 0,
       "fn": 0,
-      "tn": 16,
-      "total": 34,
-      "fixtureCount": 34,
+      "tn": 18,
+      "total": 39,
+      "fixtureCount": 39,
       "coverage": 1,
       "accuracy": 1,
       "precision": 1,
@@ -198,6 +198,18 @@
       "source": "tests/quality/benchmark.mjs",
       "notes": null
     },
+    {
+      "fixture_id": "ja-ai-04-lexicon",
+      "lang": "ja",
+      "class": "ai",
+      "detector": "patina-deterministic",
+      "expected_hot": true,
+      "predicted_hot": true,
+      "correct": true,
+      "score": 1,
+      "source": "tests/quality/benchmark.mjs",
+      "notes": null
+    },
     {
       "fixture_id": "ja-nat-01",
       "lang": "ja",
@@ -234,6 +246,18 @@
       "source": "tests/quality/benchmark.mjs",
       "notes": null
     },
+    {
+      "fixture_id": "ja-nat-04-lexicon-cold",
+      "lang": "ja",
+      "class": "natural",
+      "detector": "patina-deterministic",
+      "expected_hot": false,
+      "predicted_hot": false,
+      "correct": true,
+      "score": 0,
+      "source": "tests/quality/benchmark.mjs",
+      "notes": null
+    },
     {
       "fixture_id": "ko-ai-01",
       "lang": "ko",
@@ -306,6 +330,18 @@
       "source": "tests/quality/benchmark.mjs",
       "notes": null
     },
+    {
+      "fixture_id": "ko-ai-07-ko-diagnostic",
+      "lang": "ko",
+      "class": "ai",
+      "detector": "patina-deterministic",
+      "expected_hot": true,
+      "predicted_hot": true,
+      "correct": true,
+      "score": 1,
+      "source": "tests/quality/benchmark.mjs",
+      "notes": null
+    },
     {
       "fixture_id": "ko-nat-01",
       "lang": "ko",
@@ -402,6 +438,18 @@
       "source": "tests/quality/benchmark.mjs",
       "notes": null
     },
+    {
+      "fixture_id": "zh-ai-04-lexicon",
+      "lang": "zh",
+      "class": "ai",
+      "detector": "patina-deterministic",
+      "expected_hot": true,
+      "predicted_hot": true,
+      "correct": true,
+      "score": 1,
+      "source": "tests/quality/benchmark.mjs",
+      "notes": null
+    },
     {
       "fixture_id": "zh-nat-01",
       "lang": "zh",
@@ -437,6 +485,18 @@
       "score": 0,
       "source": "tests/quality/benchmark.mjs",
       "notes": null
+    },
+    {
+      "fixture_id": "zh-nat-04-lexicon-cold",
+      "lang": "zh",
+      "class": "natural",
+      "detector": "patina-deterministic",
+      "expected_hot": false,
+      "predicted_hot": false,
+      "correct": true,
+      "score": 0,
+      "source": "tests/quality/benchmark.mjs",
+      "notes": null
     }
   ]
 }