v1.9.4

The citation example in NOTICE.md still said v1.6.0; it now carries the current version, and the maintainer's release sync check covers it from now on. From this version on, each release is also published as a GitHub Release with its notes, not only as a tag.

---

NOTICE.md 引用範例裡的版本號從 v1.6.0 起就沒再改過，更正為目前版本，並加進維護端的發版對帳，之後會跟著標籤一起檢查。從這一版起，每個版本在 GitHub 也發 Release（附版本紀錄），不再只有標籤。

Full history / 完整版本紀錄: README.en.md · README.md

v1.9.3

Aligns six method fixes from the 2026-09-26 to 09-28 paper line. All are in the skill documents; no code changes. fetch-refs adds two checks for short titles to the file-matching step: even past the phrase threshold, a file is downgraded to LOW-CONFIDENCE, with the reason written down, when its first page prints a different DOI, or when the title has three content words or fewer and the first author's surname appears nowhere in the first eight pages (the two-word Annotated portfolios was once matched to another paper whose title contains those two words, at phrase similarity 1.00). The report now also lists LOW-CONFIDENCE, PREVIEW-ONLY and the three kinds of unobtained items separately. In verify-citations, OpenAlex title lookups must strip commas, question marks, asterisks, exclamation marks and vertical bars as well (v1.9.1 fixed only lit_map.py), and a failed OpenAlex or Semantic Scholar query is reported as "not checked", never as "not found". paper-review adds "whose manuscript is it?": when the author reviews someone else's submission for a journal or conference, most venues' confidentiality rules forbid giving it to an AI tool, so by default Claude only hands over the commands for the local tools. The rebuttal column list for revisions.tsv was missing evidence (the template and check_response.py both have it; a table built from the old description failed every DECLINE for lack of evidence). The dead-rule paragraph in doc-regress now names all three checks the tool actually makes (guard return, unregistered, undefined call) and RULES_DEREGISTERED; the note on exemplar paragraphs in co-author now reports the A/B result and treats them as optional. The README row for dead_rule_check.py said it finds rules whose anchor text was edited away; it finds rules that return at the guard because their setting is empty, and now says so.

---

對齊 9/26–9/28 論文線的六項方法修正，都在 skill 文件層，程式不變。fetch-refs 的檔案比對補上短標題的兩道檢查：片語相似度過了門檻，只要首頁印著別的 DOI，或標題只有三個以內的實詞而第一作者姓氏在前八頁完全沒出現，就降成 LOW-CONFIDENCE 並寫出理由（只有兩個詞的〈Annotated portfolios〉曾被標題含這兩個字的別篇接走，片語相似度還是 1.00）；回報時也要分別列出 LOW-CONFIDENCE、PREVIEW-ONLY 與三類拿不到的數目。verify-citations 的 OpenAlex 標題查詢同樣要先去掉逗號、問號、星號、驚嘆號與直線（v1.9.1 只修了 lit_map.py），OpenAlex 或 Semantic Scholar 查詢失敗要報「未能查核」，不能報成查無。paper-review 新增「稿件是誰的」：替期刊或研討會審別人的稿時，多數場域的審稿保密規定不允許把稿件交給 AI 工具，預設只把本機工具的指令交給使用者自己跑。rebuttal 修訂表的欄位說明漏了 evidence（範本與 check_response.py 都有這一欄，照說明建表的話，每一條 DECLINE 都會被判沒有依據）。doc-regress 的死規則說明補上從 guard return 離開、未註冊、呼叫不存在函式三種判定與 RULES_DEREGISTERED，與工具本身一致；co-author 的語感參照段落依 A/B 結果改寫成可附、非必要。README 工具表 dead_rule_check.py 那一列原本寫成抓「錨點文字改掉了」的規則，它實際抓的是設定為空、一進來就 return 的規則，一併更正。

Full history / 完整版本紀錄: README.en.md · README.md