|
42 | 42 |
|
43 | 43 | > ⚠️ 本專案與 Anthropic **無官方關聯**。部分技能需搭配外部工具或**付費資料庫(如 TEJ)**才能發揮完整功能。 |
44 | 44 |
|
| 45 | +### 🆕 最新更新(v0.13.0 · 2026-08) |
| 46 | + |
| 47 | +本版新增兩道**學術誠信防線**——都是「不做會出事、做了沒人看得見」的那種功能。 |
| 48 | + |
| 49 | +- **撤稿查核,在建檔當下就攔截(`literature-matrix-builder`)。** 引用到已撤稿的論文是實質學術風險,但既有的引用工具只查「文獻存不存在」「格式對不對」,**不查「這篇是不是已經被撤稿」**。本版讓每次 `add` 自動反查 Crossref 的撤稿通知,在文獻進庫的當下就攔下來——而不是等寫進文獻回顧才發現。 |
| 50 | + - 已用**陽性/陰性對照**實測:已知被撤論文測得出,三篇正常論文皆不誤報。只驗前者會做出一個到處亂報的工具。 |
| 51 | + - **查核失敗 ≠ 沒被撤稿**:工具標「⚠️ 查核失敗:原因」而非「未發現」。查不到與確認沒事是兩回事,這個區別攸關風險判斷。 |
| 52 | + - 實務細節:Elsevier 慣例是把撤稿通知放在**同一個 DOI**(原文被替換成撤稿聲明),若照直覺去找「另一個 DOI」會查不到。 |
| 53 | + |
| 54 | +- **雙盲投稿的身分資訊清除(`thesis-consistency-audit`)。** `.docx`/`.pptx`/`.xlsx` 是壓縮檔,作者身分藏在**五個肉眼看不到的地方**,而 Word 內建的「檢查文件」**不一定清得掉後兩項**——這是 desk reject 的常見原因: |
| 55 | + 1. `docProps/core.xml` 建立者、最後修改者 |
| 56 | + 2. `docProps/app.xml` Company(學校)、Manager(指導教授) |
| 57 | + 3. `docProps/custom.xml` 自訂屬性(常含計畫編號) |
| 58 | + 4. `word/comments.xml` **每則註解都帶作者姓名** |
| 59 | + 5. 追蹤修訂 每個 `w:ins`/`w:del` 都帶 `w:author` 與時間戳 |
| 60 | + - **紀律**:註解與追蹤修訂**含有內容**、不只是中繼資料,故預設**只報告不刪除**,要處理必須明確加旗標;清除前一律自動備份。追蹤修訂採「作者匿名化」而非刪除,避免改變文件內容。 |
| 61 | + - **誠實邊界**:工具只處理檔案中繼資料。**正文自我引用、致謝、基金計畫編號、檔名含姓名——這些才是最常見的雙盲破功點,必須人工檢查。** |
| 62 | + |
| 63 | + |
45 | 64 | ### 🆕 最新更新(v0.12.0 · 2026-08) |
46 | 65 |
|
47 | 66 | - **v0.12.0 — 台灣代碼陷阱:第二例,並升級為通則(`global-opendata-scout`)。** 繼 World Bank/OECD **完全沒有台灣資料**之後,本版實測確認第二個體系的處理方式不同但同樣會坑人:**UN Comtrade/WITS 把台灣併入代碼 `490`「Other Asia, not elsewhere specified」**。實測結果——查 `490` 回傳 218 筆,查常被誤用的 `158` 回傳 **0 筆卻不報錯**。查錯代碼會靜默得到空結果,極易誤判成「這個資料源沒有台灣資料」而放棄整個來源。本版把它從個案寫成**通則**:遇到任何聯合國體系下的國際資料庫,預設假設「台灣不會用標準 ISO 碼出現」,先查該庫怎麼處理台灣再決定研究設計,並列出三種常見處理型態(完全沒有/併入其他代碼/用非標準名稱如 "Chinese Taipei")。使用 490 時論文須揭露的三件事也一併寫明。 |
@@ -176,6 +195,25 @@ If a skill here caught a hallucinated citation, saved you an afternoon of data-w |
176 | 195 |
|
177 | 196 | > ⚠️ **Not affiliated with Anthropic.** Some skills require external tools or a **paid database (e.g., TEJ)** for full functionality. |
178 | 197 |
|
| 198 | +### 🆕 What's new (v0.13.0 · 2026-08) |
| 199 | + |
| 200 | +This release adds two **research-integrity guardrails** — the kind of feature nobody notices until its absence costs you something. |
| 201 | + |
| 202 | +- **Retraction checking at ingestion time (`literature-matrix-builder`).** Citing a retracted paper is a real academic risk, yet existing citation tools check whether a reference *exists* and whether its *format* is right — **not whether it has been retracted**. Every `add` now queries Crossref for retraction notices and flags the paper as it enters your library, rather than after it has already reached your literature review. |
| 203 | + - Verified with **both positive and negative controls**: a known-retracted paper is detected; three ordinary papers produce no false alarm. Testing only the positive case would have shipped a tool that cries wolf. |
| 204 | + - **A failed check is not a clean bill of health**: the tool reports "⚠️ check failed: reason" rather than "none found." Not knowing and knowing-it's-fine are different states, and the difference matters for risk. |
| 205 | + - Practical detail: Elsevier's convention is to place the retraction notice at the **same DOI** (the original article is replaced by the notice), so looking for "a different DOI" finds nothing. |
| 206 | + |
| 207 | +- **Blinding your submission files (`thesis-consistency-audit`).** `.docx`/`.pptx`/`.xlsx` files are archives, and author identity hides in **five places you cannot see** — and Word's own "Inspect Document" does **not** reliably clear the last two. This is a common cause of desk rejection: |
| 208 | + 1. `docProps/core.xml` — creator, last-modified-by |
| 209 | + 2. `docProps/app.xml` — Company (your university), Manager (your supervisor) |
| 210 | + 3. `docProps/custom.xml` — custom properties, often a grant number |
| 211 | + 4. `word/comments.xml` — **every comment carries its author's name** |
| 212 | + 5. Tracked changes — every `w:ins`/`w:del` carries `w:author` and a timestamp |
| 213 | + - **Discipline**: comments and tracked changes contain *content*, not just metadata, so the tool **reports without deleting** by default; removal requires an explicit flag, and a backup is always made first. Tracked-change authors are anonymised rather than stripped, so the document itself is unchanged. |
| 214 | + - **Honest boundary**: the tool handles file metadata only. **Self-citation in the body text, acknowledgements, grant numbers, and your name in the filename are the far more common ways blinding fails — those still need a human pass.** |
| 215 | + |
| 216 | + |
179 | 217 | ### 🆕 What's new (v0.12.0 · 2026-08) |
180 | 218 |
|
181 | 219 | - **v0.12.0 — The Taiwan country-code trap: a second case, generalised into a rule (`global-opendata-scout`).** After confirming that World Bank and OECD **contain no Taiwan data at all**, this release verifies a second international system that handles Taiwan differently — and just as treacherously: **UN Comtrade / WITS folds Taiwan into code `490`, "Other Asia, not elsewhere specified."** Verified live: querying `490` returns 218 records; querying `158` (a code often mistaken for Taiwan) returns **0 records without raising an error**. A wrong code yields a silent empty result, which is very easily misread as "this source has no Taiwan data" — leading researchers to abandon an otherwise usable dataset. This release turns the observation into a **general rule**: for any UN-system international database, assume by default that Taiwan will *not* appear under a standard ISO code; check how that particular database handles Taiwan *before* designing around it. Three common patterns are documented (absent entirely / folded into another code / listed under a non-standard name such as "Chinese Taipei"), along with the three things a paper must disclose when using code 490. |
|
0 commit comments