Advance tools to version 1.0.0.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

CLAUDE.md

@@ -1,68 +0,0 @@
-# Project Conventions
-
-## Analysis pipeline
-
-- LLM analysis runs via Python scripts calling the Anthropic
-  Messages API, Batch API where possible.  Steps 1 and 3 run
-  on `claude-sonnet-4-6` with `temperature=0` and thinking
-  disabled; steps 4 and 5 run on `claude-fable-5`, which
-  accepts neither parameter -- step 4 absorbs its sampling
-  variance by the majority vote, step 5 by consolidating the
-  three readings.
-- Prompt definition files live in
-  `prompts/<step>-<substep>-<task>.md` (e.g. 01-tag.md; no
-  version suffix -- versions live in git history) and are
-  passed verbatim as the system prompt.  The number names a
-  step of the research procedure, not the file: the
-  deterministic vocabulary step (step 2) has no definition
-  file yet holds its own number.  Zero padding is for
-  sorting only -- prose says "step 1", "step 3".
-- Itemwise LLM judgments (per-song coding in step 3,
-  per-keyword group selection in step 4) run the same
-  definition file three times, independently, over the same
-  input; a deterministic tally then assigns an item (a
-  (song, keyword) or (group, keyword) pair) when at least
-  two of the three runs assign it ("3 runs + majority
-  vote").  Free-generation steps run
-  twice and both outputs are pooled.  The step-5 qualitative
-  readings are neither: three independent readings per song,
-  consolidated per song and synthesized across songs by
-  their own definition files -- a qualitative protocol, not
-  a vote (see docs/methodology.md).  The vocabulary is
-  built by a deterministic subcommand (embedding +
-  clustering), not by an LLM.  If a validation outcome is
-  unexpected, revise the definition file and repeat that
-  cycle; never patch results by hand.
-- Each run of a step is archived self-contained under the
-  destination directory given explicitly on the `run-llm`
-  command line (by convention `runs/<step>/run<N>/`):
-  prompt snapshot, raw output, and `meta.json` (model ID,
-  parameters, timestamps, batch ID).  The runs of a step are
-  that many separate invocations of `run-llm`.  Replacing
-  an existing run archive requires an explicit flag;
-  superseded runs live in git history.  Deterministic steps
-  archive under `runs/<step>/` with no `run<N>` level.
-- Token usage and cost of every `run-llm` execution are
-  recorded in `docs/run-costs.md` in the same commit as the
-  run archive.
-- Scripts read the API key from the `ANTHROPIC_API_KEY`
-  environment variable (`.env`, gitignored).
-
-## Data rules
-
-- `data/source/` holds the immutable hand-placed raw files;
-  `data/captures/` is written only by the fetch commands and the
-  private import script; `data/manual/` is written only by the
-  user's own hand; `data/derived/` is written only by the
-  `build-db` subcommand.
-- Full lyrics are copyrighted: they stay in `data/captures/lyrics/`
-  (gitignored) and must never be committed or reproduced in
-  full anywhere in the repo.
-
-## Documents
-
-- `results/` holds the final tallied tables (what the paper
-  cites); `runs/` holds raw audit records.  The paper cites
-  `results/` only.
-- Any change to a definition file, the codebook, or the plan is
-  recorded in `docs/decision-log.md` with date and reason.

README.md

@@ -1,39 +1,26 @@
-# 流行音樂中「女性力量」語彙的挪用與污染——以 Billboard Year-End Hot 100（2018–2025）為例的內容分析。
+# 流行音樂中「女性力量」語彙的挪用與污染——以 Billboard Year-End Hot 100（2016-2025）為例的內容分析
 
-流行音樂中「女性力量」語彙的挪用與污染——以 Billboard Year-End
+這是研究論文《流行音樂中「女性力量」語彙的挪用與污染——以 Billboard Year-End Hot 100（2016-2025）為例的內容分析》的專案資料，包括論文本身、文件、紀錄、資料、工具程式、LLM提示詞，等等。
-Hot 100（2016–2025）為例的內容分析。
 
-台灣女性學學會 2026 年會論文之研究資料與分析程式。
+## 論文和摘要
 
-## 目錄結構
+研究論文全文和摘要，請參閱 paper/ 資料夾。
 
-見 `docs/project-structure.md`。研究步驟規劃見
+## 文件說明
-`docs/research-plan.md`；方法細節見 `docs/methodology.md`；
-人工編碼手冊見 `docs/codebook.md`；決策日誌見
-`docs/decision-log.md`。
 
-## 重現方式
+研究方法、記錄等文件，請參閱 docs/ 資料夾。
 
-1. 準備 Python 3.14+ 環境，安裝分析管線套件：
+## LLM提示
-   `pip install -e tools/`。
-2. 自 `tools/.env.example` 建立 `tools/.env`，寫入
-   Anthropic API 金鑰。
-3. 依 `docs/research-plan.md` 的階段順序，於 `tools/` 目錄下
-   執行子命令，必要輸入以位置引數、選擇性輸入以選項給定（如
-   `pop-fem-audit-tools build-db
-   ../data/source/yearend_hot100_2016_2025.csv
-   ../data/derived
-   --lyrics-dir ../data/captures/lyrics
-   --wikidata-csv ../data/captures/artists-wikidata.csv`）。
-   LLM 步驟使用 `claude-sonnet-4-6`、temperature=0、
-   thinking 關閉；輸出可逐項比對的步驟獨立執行三次，由
-   程式取三票多數決定案（「三次執行＋多數決」協定），
-   自由生成步驟兩次執行進池。
-4. 每次執行的完整紀錄（定義檔快照、原始輸出、參數）存於
-   `runs/`，可逐筆稽核。論文引用的最終資料表在 `results/`。
 
-注意：歌詞受版權保護，`data/captures/lyrics/` 不隨 repo 發布，須自行
+LLM提示請參閱 prompts/ 資料夾。
-以 `tools/` 中的抓取程式重建。
+
+## 輔助工具程式
+
+輔助工具程式請參閱 tools/ 資料夾。
+
+## 研究資料
+
+研究資料請參閱 data/ 資料夾。
 
 ## 授權
 
@@ -44,3 +31,9 @@ Hot 100（2016–2025）為例的內容分析。
 排除物件：
 1. 本專案引用之第三方歌詞，相應權利歸各該權利人所有；
 2. tools/ 資料夾內之工具程式，依該工具程式授權條款提供。
+
+## 作者
+
+楊士青<br>
+imacat@mail.imacat.idv.tw<br>
+2026/8/18<br>

docs/conventions.md

@@ -0,0 +1,69 @@
+# 常設工作規範
+
+本專案的常設工作規範。原專案 `CLAUDE.md`；移置於此，
+使 Claude Code subagent 不將其繼承入 context（盲判型
+agent 不得見之）。凡於本專案工作的主會話，動手管線、
+資料或文件之前，先讀本檔。
+
+## 分析管線
+
+管線已跑畢；本規範適用於任何重跑或擴充。
+
+- LLM 分析以 Python 腳本呼叫 Anthropic Messages API
+  執行，能用 Batch API 處即用之。步驟 1 與步驟 3 以
+  `claude-sonnet-4-6` 執行，`temperature=0`、thinking
+  停用；步驟 4 與步驟 5 以 `claude-fable-5` 執行，該
+  模型兩個參數皆不受理——步驟 4 的取樣變異由多數決
+  吸收，步驟 5 由整合三份閱讀吸收。
+- 定義檔置於 `prompts/<步><次步>-<task>.md`（如
+  1-tag.md、5a-read.md；次步以字母標示，與論文正文的
+  步驟編號一致；不帶版本號——版本即 git 歷史），逐字
+  作為 system prompt。
+- 逐項的 LLM 判斷（步驟 3 的逐首編碼、步驟 4 的逐碼
+  入群判斷）以同一份定義檔、同一份輸入獨立執行三次；
+  再由確定性計票將三次執行中至少兩次指派的項目（一個
+  （歌，關鍵字）或（群，關鍵字）配對）收入定案
+  （「三次執行＋多數決」）。自由生成步驟執行兩次，
+  兩份輸出進池。步驟 5 的質性閱讀兩者皆非：逐首三次
+  獨立閱讀，由各自的定義檔逐首整合、跨首統整——質性
+  協定，不是投票（見 docs/methodology.md）。詞彙表由
+  確定性子命令建構（嵌入＋分群），不經 LLM。驗證結果
+  不符預期時，修訂定義檔並重複該循環；絕不手改結果。
+- 一步的每次執行皆自我完備歸檔於 `run-llm` 命令列上
+  明示指定的目的目錄下（慣例為
+  `data/runs/<步驟>/run<N>/`）：定義檔快照、原始輸出與
+  `meta.json`（model ID、參數、時間戳、batch ID）。
+  一步的 N 次執行即 N 次各自的 `run-llm` 呼叫。覆蓋
+  既有執行歸檔須明示旗標；被取代的執行留在 git 歷史。
+  確定性步驟歸檔於 `data/runs/<步驟>/`，不分 `run<N>` 層。
+- 每次 `run-llm` 執行的 token 用量與費用記入
+  `docs/run-costs.md`，與執行歸檔同一 commit。
+- 腳本自環境變數 `ANTHROPIC_API_KEY` 讀取 API key
+  （`.env`，gitignored）。
+
+## 資料規則
+
+- `data/source/` 存手放後不動的原始檔；
+  `data/captures/` 只由 fetch 命令與私人匯入腳本
+  寫入；`data/manual/` 只由研究者親手寫入；
+  `data/derived/` 只由 `build-db` 子命令寫入；
+  `data/runs/` 存 LLM 執行的原始歸檔，只由執行程序
+  寫入；`data/results/` 存論文引用的定案表，只由
+  計票程序寫入。
+- 歌詞全文有版權：一律置於 `data/captures/lyrics/`
+  （gitignored），絕不 commit，亦絕不於 repo 任何處
+  全文重現。
+
+## 文件
+
+- `data/results/` 存計票後的定案表（論文所引）；
+  `data/runs/` 存原始稽核紀錄。論文只引 `data/results/`。
+- **Commit 判準**：凡能由「committed 的輸入＋committed
+  的程式」決定性再生者不 commit；凡不能者一律以文字
+  格式 commit，格式跟著上文「資料規則」一節所定的
+  層次走。例外：`data/results/` 定案表與
+  `data/derived/` 人讀報表雖可再生仍 commit——理由是
+  引用穩定性、審稿人零門檻、撰稿期數字變動可 diff；
+  兩者皆與工作儲存同一動作產出，稽核鏈無中間空缺。
+- 凡定義檔或研究規劃之更動，皆記入
+  `docs/decision-log.md`，註明日期與原因。

docs/decision-log.md

@@ -99,8 +99,8 @@
 
 ## 2026-08-02
 
-- **`import-lyrics` 不設為子命令，pilot 歌詞改以 `excludes/`
+- **`import-lyrics` 不設為子命令，pilot 歌詞改以私人腳本
-  私人腳本匯入**。理由：pilot 捕捉檔不隨論文發布，子命令
+  匯入**。理由：pilot 捕捉檔不隨論文發布，子命令
   形式會在發布的 CLI 裡留下讀者無法執行的死命令——要交待的
   是「沿用 pilot 捕捉」的事實（記於 lyrics-provenance.csv 與
   論文方法節），不是工具本身；工具移出專案，發布管線即
@@ -178,8 +178,8 @@
   來源公開可稽核，快照乾淨、overrides 縮小；論文方法節揭露
   「捕捉前經研究者查證補完」。性別以公開自我認同為準，
   推測不確定且無佐證者列疑慮清單；查證屬資料策展而非
-  分析，以 Claude Code 輔助、不走分析 API；QS 批次存
+  分析，以 Claude Code 輔助、不走分析 API；QS 批次以
-  `excludes/`（私人工作檔，不隨論文發布）。
+  私人工作檔留存，不隨論文發布。
 
 ## 2026-08-04
 
@@ -227,7 +227,7 @@
 - **Markdown 檔名一律以 dash 連接**（decision-log.md、
   research-plan.md、project-structure.md 等；與 data/ 層
   CSV 檔名慣例一致），定義檔命名慣例同步改為
-  `prompts/<task>-v<N>.md`（如 screen-v1.md），`excludes/`
+  `prompts/<task>-v<N>.md`（如 screen-v1.md），版本庫外的
   私人工作檔一併改名；`conference_abstract.md` 改名並搬入
   `paper/`——它是本次年會實際送出的摘要，與全文同屬投稿
   血脈，不是 `docs/` 的內部工作文件。全 repo 指涉同步更新。
@@ -583,8 +583,8 @@
   Billboard 署名 Pinkfong 為品牌（Wikidata Q55735607，型態
   brand），非演唱者；其唯一上榜曲 Baby Shark（2019#75）
   的實際演唱者為 Hope Segoine（KTVB 2019-03-06 報導、
-  Songfacts、經紀簡介，出處詳
+  Songfacts、經紀簡介，出處詳私人留存的
-  `excludes/quickstatements/hope-segoine.md`）。藉既有的
+  QuickStatements 草稿）。藉既有的
   署名正規化機制（`ArtistImporter.CANONICAL_ARTIST_NAMES`）
   指認：歌曲的署名字串維持榜單所印的 "Pinkfong"，解析出
   的演出者實體為 Hope Segoine。品牌無性別可言，人有；
@@ -637,10 +637,9 @@
   （sentence-transformers/all-mpnet-base-v2, revision
   e8c3b32e）對 `runs/02-cluster/source-keywords.txt` 的
   5,999 個關鍵字重新計算。組內一致性定義為組內成員兩兩
-  餘弦相似度之平均。k=50 直接取自留存的分群結果
+  餘弦相似度之平均。k=50 直接取自私人留存的分群結果；
-  （`excludes/k50/02-cluster/groups.csv`）；k=30 以同參數
+  k=30 以同參數重跑（確定性演算法），重算所得最大組恰為
-  重跑（確定性演算法），重算所得最大組恰為 491 詞，與
+  491 詞，與當時粗算紀錄吻合，佐證重算與原實驗一致。數表：
-  當時粗算紀錄吻合，佐證重算與原實驗一致。數表：
 
   | | k=30 | k=50 | k=100 |
   |---|---|---|---|
@@ -703,8 +702,8 @@
   力量語意即入選,women-power 群 11 碼,並產詞彙表外
   幻覺碼一筆),claude-fable-5 讀為詞彙化概念(要求女性
   標記,women-power 群 2 碼),與草稿盲選及深度閱讀輔助
-  判讀同讀法;sonnet 對照執行歸檔備份於
+  判讀同讀法;sonnet 對照執行歸檔另行私人備份,不入
-  `excludes/experiments/`,支出留帳。claude-fable-5 不
+  版本庫,支出留帳。claude-fable-5 不
   受理 temperature 與 thinking 參數(均不送出),無法釘
   temperature=0;執行間變異實測存在(run1/run2 於陽剛、
   脆弱邊緣碼分歧),由三票多數決吸收;詞彙表外輸出項
@@ -751,9 +750,9 @@
   編碼、票數),納入重建摘要與清空範圍,供群層次查詢;
   重建命令自此帶 `--groups results/groups.csv`。
 
-- **Fable 5 輔助判讀實驗總錄(非正式編碼;結果檔於
+- **Fable 5 輔助判讀實驗總錄(非正式編碼;結果檔私人
-  gitignored 之 `excludes/`,本條為其版本庫內的程序
+  留存、不入版本庫,本條為其版本庫內的程序錨點)**:
-  錨點)**:2026-08-08 起以 Claude Code subagent(模型
+  2026-08-08 起以 Claude Code subagent(模型
   Fable 5)進行七項深度閱讀實驗。共同程序:每首歌一個
   獨立會話、提示最少化且逐字統一、判讀者互不知情、不經
   三票制;定位為研究者的輔助判讀(初篩),不改動任何
@@ -783,7 +782,7 @@
   `male-mixed-wp-fe-feminist-reading.md` 與同名 .ods)。
   此系列即論文方法節「最後編碼的結果,再由研究者與 LLM
   (Claude Code)輔助判讀」之所指;結果檔含大量歌詞引文,
-  依著作權紀律留置 `excludes/`,不入版本庫。
+  依著作權紀律私人留置,不入版本庫。
 
 - **步驟 5:女性主義問題之質性深讀(設計定案)**:論文
   題目「挪用與污染」需要框架層的系統性證據;47 首
@@ -899,4 +898,49 @@
   performer_gender 手工修正)、四份 results 定案表、時程
   展延一至二日與剩餘撰寫工作。仍然有效的原則(先導不
   比較、提示只定格式、歌手背景防火牆、commit 判準、
-  版權規則、positionality)保留。
+  版權規則)保留。
+
+## 2026-08-17
+
+- **先導研究的地位另立常設紀錄 `docs/pilot-study.md`**:論文
+  交件後重讀投稿摘要,認清先導研究係由當時協作的 Claude Code
+  設計並執行,研究者檢視的是研究問題與答案,未逐步檢視中間
+  過程;摘要中的具體數字未經稽核,亦無歸檔可回溯。此地位
+  先前僅零星散見於本日誌各條(歌詞沿用、women-power 來歷
+  考據、粒度繼承、簿記容量),投稿摘要本身則只是產物,不足以
+  記述過程。裁定:另立 `docs/pilot-study.md`,完整記述先導
+  研究的執行方式、沿用與棄用的成果,以及它如何促成正式研究
+  的可稽核設計;投稿摘要不於樹中另存副本,其內容留在 git
+  歷史(`paper/abstract.md` 的前身)。
+
+## 2026-08-18
+
+- **工序編號改與論文正文一致**:論文正文將五個步驟的細分
+  工序以字母標示(表三「步驟3a」、表四「步驟5d」),repo
+  的檔名與歸檔目錄卻沿用 `05-01`、`05-02` 這種數字次步,
+  兩邊對不上。裁定:改為 `1-tag`、`2-cluster`、`3a-code`、
+  `4-group`、`5a-read`、`5b-consolidate`、`5c-synthesize`、
+  `5d-annotate`(`prompts/` 與 `runs/` 同步,以 `git mv`
+  保留歷史),補零一併取消。`runs/*/meta.json`
+  內記的 `prompt_path` 不追改——它是執行當下的實況記錄;
+  本日誌的既有條目同理,維持當時的名稱。`run-costs.md` 則
+  相反:它的「步驟」欄是查閱歸檔的索引,故一律換為現行
+  編號(`01-01-01-tag`→`1-tag`、`03-01-code` 與
+  `03-code`→`3a-code`),並新增末欄「原步驟名」保留執行
+  當時的名稱——索引要指得到現在的歸檔,紀錄要留得住當時
+  的形狀,兩者以兩欄並存解決。已刪除的工序(步驟 2 的
+  LLM 合併嘗試 `01-02-01-merge`、步驟 3 的仲裁
+  `03-02-arbitration`)其歸檔已不存在,無現行編號可指,
+  不另發新名,步驟欄逕標「已廢棄」,原名存於末欄。
+
+- **CLAUDE.md 移為 `docs/conventions.md`**:專案 CLAUDE.md 會
+  全文自動注入每一個 Claude Code subagent 的 context(2026-07-30
+  實測條),步驟 6 的 883 個盲判 agent 因此都看到了工作規範。
+  規範本身無研究語意(無 women-power、無假說),污染有限,但
+  「防脈絡污染」的原則應貫徹到自家工具鏈:docs/ 底下的檔案
+  不會自動注入,搬移後 subagent 除非主動翻閱否則不會看到,
+  盲判 agent 的定義檔並明令不得查閱專案材料。代價是主會話
+  也不再自動載入,須自行先讀 `docs/conventions.md`。內容
+  順帶微調:註明管線已完成、規範適用於重跑;刪去已不存在
+  的 codebook 一詞。進論文的數據一律走 Anthropic API(無
+  檔案系統,機制上不可能讀到),此層防線不受本次搬移影響。

docs/methodology.md

@@ -1,12 +1,12 @@
 # 方法細節
 
 （全文方法節底稿。演算法在執行前寫定；任何修訂記入
-`decision-log.md`。定義檔全文見 `prompts/`，執行紀錄見
+`decision-log.md`。）
-`runs/`。）
 
 ## 自然編碼管線總覽
 
-五個步驟：步驟 1 自由標註（兩次執行進池）→ 步驟 2 詞彙表
+五個步驟：步驟 1 自由標註（`claude-sonnet-4-6`，
+temperature=0、thinking 關閉，兩次執行進池）→ 步驟 2 詞彙表
 建構（詞向量分群，確定性）→ 步驟 3 全量編碼（三次執行＋
 多數決）→ 步驟 4 語意編碼群（三次執行＋多數決）→ 步驟 5
 女性主義問題之質性深讀（三次閱讀＋逐首整合＋樣態統整）。
@@ -14,13 +14,6 @@
 不接觸歌詞，步驟 2 亦不呼叫 LLM。設計原則見
 `research-plan.md`；本檔記載可重現的演算法細節。
 
-編號的所指為**研究程序的工序**，不是定義檔：步驟 1、
-步驟 3、步驟 4 與步驟 5 有定義檔（`prompts/`；步驟 5 依
-子工序有三份），步驟 2 沒有——它是單一確定性計算，由
-`cluster-keywords` 一個子命令完成。有無定義檔的區別即
-「該步是否為 LLM 判斷」，由 `prompts/` 是否存在同號檔案
-直接可見。
-
 ## 步驟 2 詞彙表建構——詞向量分群
 
 詞彙表由確定性程序產生，不經 LLM。完整分割（每個關鍵字
@@ -30,8 +23,7 @@
 ### 進池
 
 兩次標註執行的全部關鍵字取聯集、逐字串精確去重、字典序
-排列。失敗與拒答的記錄跳過（其歌曲不貢獻關鍵字）；解析
+排列。失敗與拒答的記錄跳過（其歌曲不貢獻關鍵字）。
-時偵測重複鍵，違規即失敗。
 
 ### 分群
 
@@ -39,14 +31,10 @@
   （釘定 revision），關鍵字的連字號先還原為空格再編碼，
   輸出 768 維向量並 L2 正規化。
 - **分群**：階層式聚合分群（Ward linkage），k=100。
-  向量既已正規化，歐氏距離與餘弦相似度單調對應；三種
+  向量既已正規化，歐氏距離與餘弦相似度單調對應。
-  linkage 實測比較，Ward 於各個 k 的組內一致性均最高
-  （average 與 complete 皆產生吞噬半數語料的巨大異質
-  組）。
 - **組數的取捨**：k 太小則壓縮比過高，樹上層被迫併入
-  不相干的詞，組雖大而無主題（k=30 最大組 491 詞、
+  不相干的詞，組雖大而無主題；k 太大則人工難以通覽。
-  組內一致性 0.41，成員橫跨籃球、海灘、外星人綁架）；
+  定於 100，理由是實測顯示雜物櫃
-  k 太大則人工難以通覽。定於 100，理由是實測顯示雜物櫃
   組於此始裂解為有主題的組，且編碼實測未見碼數過多的
   副作用——全量三次執行下 101 個碼全數用到；模型另行
   造出的碼共 13 筆，佔 44,149 筆標籤指派的 0.03%。
@@ -55,39 +43,23 @@
   自己產出過的關鍵字，非任何人事後撰寫。已知限制：
   組越大越異質時，medoid 只是折衷詞，可能代表不了組內
   內容（實測 `mutual-individuality` 組內一致 0.70 而
-  編碼從未使用）；此類碼於結果中呈現為零使用，據實
+  編碼從未使用）；此類碼於結果中呈現為零使用。
-  報告，不事後改名。
+- **取捨紀錄**：詞彙表分群曾比較的替代法與棄用理由，見
-- **取捨紀錄**：曾以 LLM 單發收斂（merge／cap 兩步）
+  `decision-log.md` 2026-08-05 條。
-  實作本步，四種模型六次執行全部無法維持完整分割，
+- **產物**：分群輸出中，只有實際交給模型的碼表——組名
-  已棄用（詳見 `decision-log.md` 2026-08-05；棄用的
+  加上先驗主題詞——含研究者的介入；其餘皆為分群過程本身
-  定義檔止於 git 歷史，見 `git log -- prompts/`）。
+  的機械紀錄。
-- **產物**：五份，前綴分別標示來源與結果。
-  `source-keywords.txt`（進池後的關鍵字，一行一個）
-  記錄進來的是什麼；`result-keywords.txt`（組名，一行
-  一個）與 `groups.csv`（欄位 Group、Keyword，一列一個
-  成員）記錄算出來的分割；`keywords-to-merge.json`
-  （`{"keywords": [...]}`）是實際交給模型的碼，即組名
-  加上先驗主題詞——五份中只有這一份含研究者的介入。
-  `meta.json` 記錄執行本身：進池的兩份執行歸檔與其有效
-  筆數、嵌入模型與釘定 revision、分群參數與組數、外加
-  的先驗詞、關鍵字總數，以及產生數字的套件版本。凡命令
-  列上的選擇與環境事實皆在此，不記時間戳與輸入雜湊
-  ——前者使同環境重跑逐位元組可再生，後者只會重述 git
-  已保證的事。
 - **可重現性**：同一輸入、同一釘定模型、同一參數逐次
   重現。不同 CPU／BLAS 實作的浮點尾數差異可能使邊界
-  詞的歸屬翻動，屬已揭露的限制；論文所用碼表逐字
+  詞的歸屬翻動，屬已揭露的限制。
-  commit，引用單位為該份定案檔案。
 
 ### women-power 的注入
 
 定案詞彙表為 100 個分群組名再加上 `women-power` 一詞，
 共 101 個碼。`women-power` 是研究者任意決定的先驗主題（即本
 論文的主題本身），不由資料產生，屬揭露的儀器介入。該詞
-於執行時以 `--extra-keyword` 明示加入，不寫死在程式裏
+的加入於每次執行皆為明示、可稽核的介入，而非無聲內建於
-——研究者的介入因此每次都出現在重現命令上，而非無聲
+判斷邏輯；分群結果本身不含它，只有交給模型的碼表含它。
-發生；分群結果的兩份產物不含它，只有交給模型的碼表含
-它。
 
 注入而非另設篩選軌的理由：讓研究者的主題詞與模型自己
 收斂出的類別（分群已自行長出 `female-empowerment` 等組）
@@ -97,24 +69,17 @@
 
 ## 步驟 3 編碼的三次執行與多數決
 
+- **模型**：`claude-sonnet-4-6`，temperature=0、thinking 關閉。
 - **三次執行**：同一份定義檔、同一份輸入檔，獨立執行
-  三次，三份歸檔並列（`runs/03-code/run1`、`run2`、
+  三次，三份歸檔並列，彼此無先後主從之別。
-  `run3`），彼此無先後主從之別。
 - **多數決**：一首歌的一個標籤，三次執行中至少兩次標出
   即收入定案編碼。三票不平手，裁決規則因此無例外條款，
-  計票由確定性子命令完成（見交接契約）。
+  計票由確定性程序完成（見交接契約）。
-- **第三票取全量**：只對前兩次分歧的標籤補問第三票，
+- **第三票取全量**：第三次執行同為全量——全部歌曲、
-  計票結果相同；仍採全量執行——全部歌曲、全部關鍵字
+  全部關鍵字——使三票在同一條件下取得。
-  ——使三票在同一條件下取得。
 - **對邊緣標籤的作用**：兩次執行只分得出「兩次皆標」與
   「僅一次標」；三次執行還分得出 3-0 與 2-1，故「定案
   編碼中有多少比例僅以一票之差成立」成為可報告的量。
-  至於判定本身，兩次執行相左的標籤在何種協定下都由第三
-  個判斷定奪，票數不使不確定性消失：某標籤於單次執行被
-  標出的傾向若恰為一半，任何票數皆為擲幣。三票之效在
-  傾向偏離一半處——多數決將判定推向該傾向本身（單次
-  0.7 者為 0.78，0.9 者為 0.97），程序重跑的一致性因而
-  高於單次執行，唯獨恰半處無從改善。
 
 ## 步驟 4 語意編碼群
 
@@ -125,103 +90,79 @@
   由 LLM 依編碼名的字面語意判斷。
 - **任務**：每筆輸入為一個群名加 101 個編碼的字母序
   清單，輸出為入選編碼的單層 JSON 陣列；定義檔
-  `prompts/04-group.md` 只定格式，不含任何群的語意定義。
+  只定格式，不含任何群的語意定義。
-- **模型**：`claude-fable-5`（步驟 1、3 為
+- **模型**：`claude-fable-5`；取樣變異由多數決吸收；
-  `claude-sonnet-4-6`）。該模型不受理 `temperature` 與
-  `thinking` 參數，兩者均不送出；取樣變異由多數決吸收。
   模型裁定的理由與對照實驗見決策日誌。
 - **三次執行**：同一份定義檔、同一份輸入檔，獨立執行
-  三次，歸檔並列（`runs/04-group/run1`、`run2`、`run3`）。
+  三次，歸檔並列。
 - **多數決**：一個（群，編碼）配對，三次執行中至少兩次
-  入選即屬該群；不在 101 碼詞彙表內的輸出項無效，每筆
+  入選即屬該群；不在 101 碼詞彙表內的輸出項無效，逐筆
-  丟棄印於標準錯誤。計票由確定性子命令 `tally-groups`
+  記錄後丟棄。計票由確定性計票程序完成，合法碼
-  完成：`tally-groups <執行歸檔 1> <執行歸檔 2> <執行歸檔
+  清單之產法同步驟 3，結果為定案分群表。
-  3> <合法碼清單> <輸出 CSV>`，合法碼清單之產法同步驟 3。
-  定案分群寫入 `results/groups.csv`，欄位 `Group`、
-  `Keyword`、`Votes`，列序先依群名、再依編碼，一律以
-  Unicode 碼位比較，換行為 CRLF。
-- **工作儲存**：`build-db --groups <定案分群 CSV>` 將定案
-  分群逐欄照存入 `groups` 資料表（群、編碼、票數），供
-  群層次查詢。
 
 ## 步驟 5 女性主義問題之質性深讀
 
-本步驟之 5-1 至 5-3 為**質性閱讀,非編碼**:輸出為自由
+本步驟之 5a 至 5c 為**質性閱讀，非編碼**：輸出為自由
-文字的問題閱讀報告,無可逐項機械比對的單位,故不適用
+文字的問題閱讀報告，無可逐項機械比對的單位，故不適用
-三票多數決與仲裁;
+三票多數決與仲裁；
-三次獨立閱讀為分析者三角檢核,逐首整合為整合而非裁決,
+三次獨立閱讀為分析者三角檢核，逐首整合為整合而非裁決，
-跨首統整之產出為草稿,終審與詮釋由研究者為之。論文引用
+跨首統整之產出為草稿，終審與詮釋由研究者為之。論文引用
 本步驟時不作次數宣稱。
 
-- **對象**:定案編碼含 `women-power` 或
+- **對象**：定案編碼含 `women-power` 或
   `female-empowerment` 的 145 首歌。
-- **5-1 逐首閱讀**:每筆輸入為一首歌的完整歌詞逐字全文,
+- **5a 逐首閱讀**：每筆輸入為一首歌的完整歌詞逐字全文，
-  不含歌名與演唱者(盲讀);定義檔
+  不含歌名與演唱者（盲讀）。同一份定義檔、同一份輸入檔，
-  `prompts/05-01-read.md`。同一份定義檔、同一份輸入檔,
+  獨立執行三次，歸檔並列。
-  獨立執行三次,歸檔並列(`runs/05-01-read/run1`、
+- **5b 逐首整合**：每筆輸入為該首歌的三份閱讀報告
-  `run2`、`run3`)。
+  （不含歌詞）；以問題機制為單位保守合併，標收斂註記
-- **5-2 逐首整合**:每筆輸入為該首歌的三份閱讀報告
+  （（3/3）、（2/3）），主清單僅列兩讀以上提出者，單讀發現
-  (不含歌詞);以問題機制為單位保守合併,標收斂註記
+  以一行存目；定義檔，執行一次。
-  ((3/3)、(2/3)),主清單僅列兩讀以上提出者,單讀發現
+- **5c 樣態統整**：單筆輸入為一批整合報告；歸納問題
-  以一行存目;定義檔 `prompts/05-02-consolidate.md`,
+  **樣態**——問題呈現與運作的重複形態，非問題分類，
-  執行一次,歸檔 `runs/05-02-consolidate/run1`。
+  代表引句僅取自主清單。四種輸入範圍各執行
-- **5-3 樣態統整**:單筆輸入為一批整合報告;歸納問題
+  一次：全 145 首之基底統整，及依演唱聲音之性別切分
-  **樣態**——問題呈現與運作的重複形態,非問題分類,
+  之三個發話脈絡統整——男聲（male）、女聲（female）、
-  代表引句僅取自主清單;定義檔
+  混合（mixed）；基底看橫貫各
-  `prompts/05-03-synthesize.md`。四種輸入範圍各執行
+  脈絡之樣態，分組看各權力脈絡下之樣態。genderfluid
-  一次:全 145 首之基底統整(歸檔
+  與 non-binary 共 3 首不設群——樣本數不支持歸納——
-  `runs/05-03-synthesize/run1`),及依 `performer_gender`
+  僅入基底統整。同一輸入不重複
-  (演唱聲音之性別)切分之三個發話脈絡統整——男聲
+  執行：自由歸納之產出無機械合併可言，其變異由 5a
-  (male,歸檔 `run2`)、女聲(female,歸檔 `run3`)、
+  三讀、5b 整合與草稿地位承接，四份草稿互為對照，由
-  混合(mixed,歸檔 `run4`);基底看橫貫各脈絡之樣態,
+  研究者終審裁決。
-  分組看各權力脈絡下之樣態。genderfluid 與 non-binary
+- **5d 樣態標註**：以（歌，樣態）對為可逐項機械比對之
-  共 3 首不設群——樣本數不支持歸納——僅入基底統整,
+  單位，回歸「三次執行＋多數決」協定。對象為「有問題」
-  由研究者以個案閱讀。同一輸入不重複執行:自由歸納之
+  的歌——三讀中至多一個「無」（多數決精神；恰兩「無」
-  產出無機械合併可言,其變異由 5-1 三讀、5-2 整合與
+  者其整合報告主清單必為空，與 5b 主清單規則自洽），
-  草稿地位承接,四份草稿互為對照,由研究者終審裁決。
+  計 111 首（男聲 12、女聲 69、混合 29、genderfluid 1）。
-- **5-4 樣態標註**:以 (歌, 樣態) 對為可逐項機械比對之
+  樣態表為三份分組統整草稿原文：男聲 13 條（M1–M13）、
-  單位,回歸「三次執行＋多數決」協定。對象為「有問題」
+  女聲 14 條（F1–F14）、混合 16 條（X1–X16）；基底 15 條
-  的歌——三讀中至多一個「無」(多數決精神;恰兩「無」
-  者其整合報告主清單必為空,與 5-2 主清單規則自洽),
-  計 111 首(男聲 12、女聲 69、混合 29、genderfluid 1)。
-  樣態表為三份分組統整草稿原文:男聲 13 條(M1–M13)、
-  女聲 14 條(F1–F14)、混合 16 條(X1–X16);基底 15 條
   不入矩陣——全體歸納之一條樣態可能疊合不同方向的
-  權力關係(男對女、女對男),實為多個樣態共用一名。
+  權力關係（男對女、女對男），實為多個樣態共用一名。
-  檢驗範圍:男聲樣態不檢驗純女聲歌、女聲樣態不檢驗
+  檢驗範圍：男聲樣態不檢驗純女聲歌、女聲樣態不檢驗
-  純男聲歌(發話位置範疇錯置,檢查無意義);混合樣態
+  純男聲歌（發話位置範疇錯置，檢查無意義）；混合樣態
-  檢驗全部(其男女聲部無系統化切分方式,無意義之標註
+  檢驗全部（其男女聲部無系統化切分方式，無意義之標註
-  容忍之);genderfluid 歌三套全查(無自身透鏡,發話
+  容忍之）；genderfluid 歌三套全查（無自身透鏡，發話
-  位置無法先驗決定),其歸屬引用維持個案地位。每筆
+  位置無法先驗決定），其歸屬引用維持個案地位。每筆
-  輸入為一首歌之整合報告(「僅單獨提及」行於組裝時
+  輸入為一首歌之整合報告（「僅單獨提及」行於組裝時
-  剝除,標註僅依主清單)與該首適用之樣態表;定義檔
+  剝除，標註僅依主清單）與該首適用之樣態表；定義檔
-  `prompts/05-04-annotate.md`,獨立執行三次
+  獨立執行三次，（歌，樣態）對
-  (`runs/05-04-annotate/run1`~`run3`),(歌, 樣態) 對
   得兩票以上者定案。
-- **模型**:`claude-fable-5`(與先導深讀同儀器;
+- **模型**：`claude-fable-5`。
-  `temperature` 與 `thinking` 參數不適用,均不送出)。
+- **輸入組裝**：各步輸入檔由確定性程序自上游產物組裝。
-- **輸入組裝**:確定性行內腳本。5-1:145 首依歌曲 ID
+  5d 之樣態條目自分組統整草稿機械切出，代表引句不
-  升序,`content` 為歌詞逐字全文;5-2:每筆
+  隨附——引句出自特定歌曲，判該曲時形同預答。
-  `{"reports": [run1 輸出, run2 輸出, run3 輸出]}`;
-  5-3:單筆以 `song-<ID>` 為鍵、整合報告為值之 JSON
-  物件,鍵集合為該次統整之範圍(基底為全 145 首,
-  分組依工作庫 `performer_gender` 切分);5-4:每筆
-  `{"report": 主清單, "patterns": [{"id", "name",
-  "description"}]}`,樣態條目自分組統整草稿機械切出,
-  代表引句不隨附——引句出自特定歌曲,判該曲時形同
-  預答。各輸入檔之 SHA-256 記入該步 meta。
 
 ## 女性力量候選集
 
-候選集為兩類歌曲的合集：定案編碼含 `women-power` 者，
+候選集為定案編碼含 `women-power` 或 `female-empowerment`
-以及定案編碼含研究者指認之女性力量概念域分群組者。
+（步驟 4 女性力量群的兩個編碼）之歌曲聯集：wp 66 首、
-指認於詞彙表定案後、黃金標準編碼開始前完成，指認清單
+fe 144 首，聯集 145 首。
-與理由記入決策日誌。
 
 ## 軌跡對映（診斷用）
 
 沿收斂軌跡的機械對映：原始關鍵字 →（兩份標註執行歸檔的
-`output.jsonl`）歌曲、原始關鍵字 →（分群）組，純程式查表，
+原始輸出）歌曲、原始關鍵字 →（分群）組，純機械查表，
 決定性。以其結果與步驟 3 直接編碼的差異率作為「收斂軌跡
 扭曲」的診斷量，不作主結果。
 
@@ -230,78 +171,36 @@
 每一步的輸出如何變成下一步的輸入，皆為確定性程序，規則
 明定如下：
 
-- **歌詞輸入檔（步驟 1）**：`export-llm-input` 自工作
+- **歌詞輸入檔（步驟 1）**：由確定性的匯出程序自工作
-  儲存產出，每筆 `{"id": "song-<ID>", "content": <歌詞>}`，
+  儲存產出，一筆一首歌，依歌曲 ID 升序。步驟 3 的輸入
-  依歌曲 ID 升序。步驟 3 的輸入由同一子命令、同一工作
+  由同一匯出程序、同一工作儲存產出（見下），兩步的語料
-  儲存產出（見下），兩步的語料同一性由此成立；各步
+  同一性由此成立。
-  輸入檔的 SHA-256 記入該步 meta。
+- **步驟 1 → 2**：確定性的分群程序讀兩份執行歸檔的
-- **步驟 1 → 2**：`cluster-keywords` 讀兩份執行歸檔的
+  執行紀錄，進池後直接分群，產出詞彙表與交給
-  `output.jsonl`（一律以換行字元 `\n` 切行——歌詞含
+  模型的碼表。
-  U+0085 等控制字元時，`str.splitlines()` 類的通用切行
+- **步驟 2 → 3 輸入檔**：同一匯出程序自工作儲存產出
-  會截斷 JSON 字串，實測踩中），進池後直接分群，一次
+  步驟 3 的輸入，一筆一首歌，兼含歌詞與定案碼表，依
-  產出上列五份檔案。
+  歌曲 ID 升序。碼表以參數傳入而非填進定義檔——定義
-- **步驟 2 → 3 輸入檔**：`export-llm-input --extras
-  <定案碼表>` 自工作儲存產出步驟 3 的輸入，每筆
-  `{"id": "song-<ID>", "content": <字串>}`，`content` 為
-  固定鍵序序列化的 `{"lyrics": …, "keywords": [...]}`，
-  依歌曲 ID 升序。碼表以參數傳入而非填進定義檔——定義
   檔只規定任務形狀，換詞彙表、換演算法都不必改它。
-- **步驟 3 定案**：`tally-codings <執行歸檔 1> <執行歸檔
+- **步驟 3 定案**：確定性的計票程序讀三份執行歸檔的
-  2> <執行歸檔 3> <輸出 CSV> --corrections <更正表>
+  執行紀錄，依序套用更正表、驗證所有標籤皆在合法碼
-  --valid-keywords <合法碼清單>` 讀三份執行歸檔的
-  `output.jsonl`，依序套用更正表、驗證所有標籤皆在合法碼
   清單之內、計票。
-- **更正表**：`data/manual/coding-corrections.csv`，研究者
+- **更正表**：研究者逐列校定的人工著作，逐筆更正標籤
-  逐列校定的人工著作，欄位 `Song ID`、`Run`、`Type`、
+  或引述——以替代字串取代，或刪去該筆標籤指派或該句
-  `To Be Replaced`、`Correct Term`。`Type` 為 `keyword`
+  引述。一筆引述更正套用於該首歌該次執行的所有出現處。
-  或 `evidence`，分別更正標籤與引述；`Correct Term` 為
+  表中任一列若在資料中找不到對應者，即中止；校定的
-  替代字串，或 `**REMOVE**` 表示刪去該筆標籤指派（`keyword`）
+  判準記於 `decision-log.md`。
-  或該句引述（`evidence`）。一筆 `evidence` 更正套用於該
+- **合法碼清單**：自詞彙表產出：分群組名加上
-  首歌該次執行的所有出現處。兩個文字欄以歌詞慣例「 / 」
+  `women-power`。
-  表示換行（與載入後的執行紀錄同一表示法，逐字比對、不再
+- **定案編碼表**：歌名與演出者名銜逐首查工作儲存取得。
-  轉換），故一列一行，純文字工具可逐列處理。表中任一列若
+  每個定案標籤隨附其在計票中各份執行所引的
-  在資料中找不到對應者，即中止；校定的判準記於
+  歌詞行，供逐碼查核（三份執行彼此無先後主從之別）。
-  `decision-log.md`。
-- **合法碼清單**：純文字、一行一個碼，自詞彙表產出：
-  `{ cat runs/02-cluster/result-keywords.txt; echo
-  women-power; } | sort`。
-- **定案表**：`results/codings.csv`，欄位 `Song`、
-  `Artist Credit`、`Keyword`、`Quote`，一列一個標籤。歌名
-  與演出者名銜逐首查工作儲存取得，故本子命令須在
-  `build-db` 之後執行。`Quote` 為該標籤在計票中各份執行
-  所引的歌詞行：各份的引述串接後逐字去重，按 Unicode
-  碼位排序，以單一 `|` 相接（三份執行彼此無先後主從之
-  別，引述之序取決於引述本身）；引述內的換行於執行紀錄
-  載入時一次換成歌詞慣例「 / 」，此後更正表、定案表與
-  工作儲存全鏈路同一表示法，不再還原。「 / 」的無歧義性
-  是語料事實而非結構保證：全 883 首歌詞經窮舉查核不含
-  「 / 」；換語料須重查。列序依印出的前三欄依序排：
-  歌名、演出者名銜、
-  標籤，一律以 Unicode 碼位比較，換行為 CRLF（同專案
-  其他 CSV）。
-- **序列化通則**：所有中間檔為 UTF-8，欄序、鍵序與元素
-  序皆依上列規則明定，無時間戳、無隨機成分；JSON 解析
-  一律偵測重複鍵，違規即失敗。人讀為主的產物採純文字或
-  CSV（CSV 依 RFC 4180，標題列字首大寫），機器交接檔採
-  JSON。給定相同的 LLM 執行輸出，全部交接產物逐位元組
-  可再生。
 
 ## 執行與稽核
 
-- LLM 步驟以 `run-llm <定義檔> <輸入檔> <歸檔目錄>`
+- LLM 步驟以批次執行程序執行，一份定義檔配一份
-  執行；一步的 N 次執行＝重現命令清單上的 N 行命令，
+  輸入檔；一步的 N 次執行為 N 次各自獨立的呼叫，各自
-  各自歸檔（`runs/<步驟>/run1`、`run2`，三票制步驟另有
+  歸檔自我完備。
-  `run3`）。
-- 確定性步驟（進池、分群、計票、對映）為子命令，其
-  輸入輸出檔同隨 `runs/` 歸檔；因無執行變異，歸檔目錄
-  下不分 `run<N>` 層。
 - Batch API 的每筆請求自含全部脈絡且互不可見（平台
   契約），歌與歌之間的獨立性由此成立；各次執行的獨立
   性由「一次呼叫、一個批次、一份歸檔」的執行結構自明。
-- 每次 `run-llm` 執行的 token 用量與費用記入
-  `run-costs.md`，被取代的執行一併保留供總支出核算。
-
-## 映射分析方法
-
-（依 2026-07-30 決策，於看到結果前寫定；待黃金標準
-編碼展開前補入。）

docs/output-validation.md

@@ -1,11 +1,11 @@
 # LLM 輸出的契約查核
 
 （2026-08-06 量測。對象為步驟 3 的三份執行歸檔
-`runs/03-code/run1`–`run3`，共 883 首歌、44,149 筆標籤
+`data/runs/3a-code/run1`–`run3`，共 883 首歌、44,149 筆標籤
 指派、44,146 句引述。歌詞以模型實際看到的那一份為準，即
 `tools/instance/llm-input-code.jsonl`。）
 
-定義檔 `prompts/03-code.md` 對輸出下了四條明確要求：只用
+定義檔 `prompts/3a-code.md` 對輸出下了四條明確要求：只用
 給定的碼且拼寫照給、每個標出的碼恰引一行歌詞、引述逐字、
 輸出為合法 JSON 且無其他文字。本檔記錄這四條各自被遵守到
 什麼程度。
@@ -46,9 +46,8 @@ song-750 run3: -use=[]  -abuse=["I got a thing for the hard
                                  liquor on ice"]
 ```
 
-模型寫錯後綴、已輸出的 token 收不回，遂以空陣列收束該鍵，
+song-750 的 run1 與 run2 則整筆使用錯拼的鍵並附上引述，
-再於正確的鍵補上引述。song-750 的 run1 與 run2 則整筆使用
+故該首的三票分裂於兩種拼寫之間。
-錯拼的鍵並附上引述，故該首的三票分裂於兩種拼寫之間。
 
 ## 引述的存在
 

docs/pilot-study.md

@@ -0,0 +1,85 @@
+# 先導研究的來歷與地位
+
+（2026-08-17 認清並記錄。本檔記述正式研究之前的先導研究:
+它做了什麼、研究者檢視到哪一層、哪些成果被沿用、哪些被
+棄用,以及它如何促成正式研究的設計。）
+
+## 一、先導研究是什麼
+
+正式研究之前,研究者曾以 Claude Code 對 Billboard Year-End
+Hot 100(2018–2025、684 首)的歌詞做過一輪探索性分析,檢視
+「女性力量」語彙的使用狀況,並附帶分析 pussy 一詞作為女性
+代稱的修辭。2026 年 4 月投出的研討會摘要即根據該輪分析撰寫。
+
+## 二、它實際的執行方式
+
+**研究問題與方向由研究者提出**:女性力量語彙被父權框架取用、
+pussy 作為代稱的跨性別使用,都是研究者長期關注的問題。
+
+**設計與執行由 Claude Code 進行**:分析管線、分類架構(genuine
+／peripheral／fake 三分類)、五種「假女性力量」類型、以及
+frame-aware 提示修正實驗,均由當時協作的 Claude Code 設計並
+執行。
+
+**研究者檢視的是問題與答案,不是中間過程**:研究者確認研究
+問題的方向正確、答案值得追究,即據以投稿;分類類目如何產生、
+提示詞如何下、數字如何算出,並未逐步檢視。投稿摘要中的具體
+數字(如「44% 屬於假女性力量」、62 首的三分類、53 首的 pussy
+修辭分類、假率由 44% 降至 5% 以下)研究者未曾稽核,其產生
+過程亦無歸檔可回溯。
+
+**當時的判斷**:投稿未上即罷;若上,正式研究由研究者自行
+設計與執行。事實上正式研究即是如此重做的。
+
+## 三、被沿用的成果
+
+- **歌詞捕捉檔**:先導研究蒐集的 lyrics.json(684 首,
+  2018–2025)以私人腳本(不入版本庫)匯入歌詞快取,只取
+  識別欄位與歌詞本文,先導的分析欄位一概不匯入;出處記於
+  `data/captures/lyrics-provenance.csv`
+  (見 `decision-log.md` 2026-07-31、08-02 條)。
+- **假說方向**:女性力量語彙的挪用與污染,成為正式研究的
+  研究問題。
+- **粒度選擇**:正式研究鎖定 thematic keywords 這一粒度,
+  係繼承先導研究三種粒度的比較結果(keywords 過碎、themes
+  過早抽象)。
+- **`women-power` 一詞的來歷**:考據先導研究的 local agent
+  存檔可知,其第一步指令含數十個範例 thematic keywords,
+  其中即有 women-power——為當時協作的 Claude Code 依研究者
+  長期表達的關注主動加入(研究者端播種,非明示指定);先導
+  的標籤 `women-power-and-empowerment` 則是第三步強制合併
+  兩個關鍵字的管線人工產物(見 `decision-log.md` 2026-08-04
+  條)。
+- **簿記容量的教訓**:先導研究九百餘詞可以在單一回應內完成
+  分組;此法未沿用,詞彙表改由確定性程序產生(見
+  `decision-log.md` 2026-08-06 條)。
+
+## 四、被棄用的成果
+
+以下先導研究的產物未進入正式研究,論文亦未引用:
+
+- **genuine／peripheral／fake 三分類與「44% 假女性力量」**:
+  構念與分母均未經稽核回溯,未沿用。
+- **五種「假女性力量」類型(A–E)**:未沿用。
+- **pussy 一詞的修辭分類**:未沿用。
+- **frame-aware(框架感知)提示修正法**:先導研究以框架判準
+  寫進提示以提高準確率;未沿用。
+- **三個獨立 LLM subagent＋人工仲裁的流程**:未沿用,原因見
+  `decision-log.md` 2026-07-30 條。
+
+## 五、它促成了可稽核的設計
+
+先導研究的根本限制不在結論對錯,而在**不可稽核**:沒有定義檔
+快照、沒有原始輸出歸檔、沒有參數紀錄,因此任何一個數字都無法
+回溯到產生它的那一次執行。正式研究的可稽核設計,正是針對這一點
+而立。
+
+換言之,正式研究對 LLM 輸出所採取的「不信任、須查核」立場,
+其第一個案例就是先導研究本身。
+
+## 六、現存的紀錄
+
+- 投稿時的先導研究摘要:已於 git 歷史中(`paper/abstract.md`
+  的前身),樹中不另存副本——摘要只是產物,不足以記述過程,
+  故另立本檔。
+- 先導研究的歌詞捕捉檔沿用事實:`data/captures/lyrics-provenance.csv`。

docs/project-structure.md

@@ -1,134 +0,0 @@
-# 專案目錄結構
-
-（2026-07-30 討論定案；2026-07-31 更新為 tools/ 子專案與
-SQLite 工作儲存架構；2026-08-17 依完成後的現況更新）
-
-```
-pop-fem-audit/
-├── README.md                  # 專案說明、重現步驟
-├── CLAUDE.md                  # 極簡工作規範（subagent 會讀到，
-│                              #   絕不放理論、codebook、預期結果）
-├── .gitignore                 # captures/lyrics/、.env、scratch
-├── data/                      # 依生命週期分層（文字格式）
-│   ├── source/                # 源頭：手放後不動
-│   │   └── yearend_hot100_2016_2025.csv   # 原始榜單
-│   ├── captures/              # 外部捕捉：只由 fetch 命令與
-│   │   │                      #   私人匯入腳本寫入
-│   │   ├── artists-wikidata.csv           # Wikidata 快照
-│   │   ├── lyrics-provenance.csv          # 歌詞出處
-│   │   └── lyrics/                        # 歌詞 .txt 快取
-│   │                                      #   （gitignored，版權）
-│   ├── manual/                # 人工著作：只由研究者手寫
-│   │   ├── coding-corrections.csv         # 編碼與引述的校對表
-│   │   └── performer-gender-corrections.csv  # 演唱聲音性別的
-│   │                                      #   手工修正
-│   └── derived/               # 衍生：只由 build-db 寫入
-│       ├── songs.csv          # 歌曲報表（人讀；進 git）
-│       └── artists.csv        # 歌手報表（人讀；進 git）
-├── prompts/                   # LLM 定義檔（逐字作為 system prompt）
-│   └── <步>-<次步>-<task>.md  # 01-tag.md、03-code.md、
-│                              #   04-group.md、05-01-read.md、
-│                              #   05-02-consolidate.md、
-│                              #   05-03-synthesize.md、
-│                              #   05-04-annotate.md
-│                              #   （步內僅一個執行時省略次步）
-│                              #   不帶版本號，版本即 git 歷史
-│                              #   （編號的所指是工序：確定性
-│                              #   的步驟 2 無定義檔仍佔一號；
-│                              #   補零只為排序）
-├── tools/                     # 輔助工具子專案（src-layout）
-│   ├── pyproject.toml         #   發行名 pop-fem-audit-tools；
-│   │                          #   pip install -e tools/ 安裝
-│   ├── README.rst  LICENSE  MANIFEST.in  .env.example  .gitignore
-│   ├── docs/                  # Sphinx API 文件
-│   ├── instance/              # SQLite 工作儲存（generated、
-│   │                          #   gitignored；含歌詞全文）
-│   ├── src/pop_fem_audit_tools/
-│   │   ├── __main__.py        # 套件 CLI 進入點（分派子命令）
-│   │   ├── commands/          # CLI 子命令模組（登記於 __init__）
-│   │   │   ├── build_db.py    # build the SQLite working store
-│   │   │   │                  #   from the inputs
-│   │   │   ├── export_llm_input.py # export the LLM input JSONL
-│   │   │   │                       #   (lyrics only) from the
-│   │   │   │                       #   working store
-│   │   │   ├── fetch_artists.py # fetch artist metadata from
-│   │   │   │                    #   Wikidata into the snapshot CSV
-│   │   │   ├── fetch_lyrics.py # fetch missing lyrics from the
-│   │   │   │                   #   public APIs into the lyrics dir
-│   │   │   ├── cluster_keywords.py # pool the tagging runs'
-│   │   │   │                       #   keywords and cluster them
-│   │   │   │                       #   into the codes (step 2)
-│   │   │   ├── tally_codings.py    # settle step 3 by majority
-│   │   │   ├── tally_groups.py     # settle step 4 by majority
-│   │   │   ├── tally_annotations.py # settle step 5-4 by majority
-│   │   │   └── run_llm.py     # API 執行器：一份定義檔＋一份輸入
-│   │   │                      #   →歸檔至指定目錄（Batch API）；
-│   │   │                      #   多次執行的計票由獨立子命令承擔
-│   │   ├── config.py          # pydantic-settings 設定（.env）
-│   │   ├── database.py        # SQLAlchemy engine / session / Base
-│   │   ├── models.py          # SQLAlchemy ORM 資料模型
-│   │   └── utils.py           # 共用工具（format_duration）
-│   └── tests/                 # 單元測試（unittest）
-├── runs/                      # 現行執行的完整稽核紀錄（進 git；
-│   │                          #   重跑同一 run 須明示 --replace）
-│   ├── <步驟名>/              #   一步一個目錄（01-tag、03-code、
-│   │                          #   04-group、05-01-read、
-│   │                          #   05-02-consolidate、
-│   │                          #   05-03-synthesize、05-04-annotate）
-│   │   └── run<N>/            #   LLM 步驟：每個 run 一份自我
-│   │       ├── prompt.md      #   完備歸檔（定義檔快照）
-│   │       ├── output.jsonl   #   該次執行原始輸出
-│   │       └── meta.json      #   model ID、temperature、時間戳、
-│   │                          #   batch ID、token 用量
-│   └── 02-cluster/           # 確定性步驟：無執行變異，
-│                              #   不分 run<N> 層
-├── results/                   # 論文引用的定案表 CSV（計票子命令
-│   │                          #   產出；「可再生仍 commit」的例外）
-│   ├── codings.csv            # 步驟 3 定案編碼
-│   ├── groups.csv             # 步驟 4 定案編碼群
-│   ├── patterns.csv           # 步驟 5-3 定案樣態表
-│   ├── annotations.csv        # 步驟 5-4 定案歌×樣態
-│   └── pattern-matrix.csv     # 前四者的人讀寬表
-├── docs/
-│   ├── research-plan.md       # 研究步驟規劃（本檔之姊妹篇）
-│   ├── project-structure.md   # 本檔
-│   ├── output-validation.md   # LLM 輸出的契約查核紀錄
-│   ├── decision-log.md        # 決策日誌：每次改定義檔的原因
-│   ├── run-costs.md           # 每次執行的 token 用量與費用
-│   └── methodology.md         # 方法細節（全文方法節底稿；
-│                              #   映射分析方法須在看結果前寫定）
-└── paper/
-    ├── abstract.md            # 摘要
-    └── 流行音樂中「女性力量」….odt  # 全文
-```
-
-## 設計理由
-
-- **`runs/` 自我完備**：每個執行目錄含定義檔快照 + 原始輸出 +
-  meta，讀者不需 git 考古即可稽核任一筆結果。
-- **`runs/`（原始稽核資料）與 `results/`（最終表）分離**：
-  論文只引 `results/`，其來源可回溯至 `runs/`。
-- **`prompts/` 檔名不帶版本號**：版本即 git 歷史，失敗的
-  版本不保留；論文引用的單位是 `runs/` 內隨執行保存的定義檔
-  快照（每個執行目錄自我完備），不需檔名可指的版本名。
-- **工作儲存的資料表**：`songs`（含 `performer_gender`＝演唱
-  聲音的性別）、`chart_entries`、`artists`、`song_artists`、
-  `codings`（定案編碼：一歌一標籤一列，`quotes` 存該標籤所據的
-  歌詞引述，多句以 `|` 相接）、`groups`（語意編碼群）、
-  `patterns`（深讀樣態）、`annotations`（歌×樣態定案矩陣）。
-  各定案表經 `build-db` 的 `--codings`、`--groups`、
-  `--patterns`、`--annotations` 匯入，性別修正經
-  `--gender-corrections` 套用，與其餘資料同一交易，儲存不會
-  半建；詳見 `research-plan.md`「資料儲存與模型」。
-- **Commit 判準**：能由「committed 輸入＋程式」決定性再生者不
-  commit（SQLite 工作儲存、LLM 輸入檔）；源頭、捕捉、人工著作
-  一律以文字 commit。「可再生仍 commit」的例外有二：
-  `results/` 報表（引用穩定性、審稿人零門檻、撰稿期可 diff）
-  與 `data/derived/` 人讀報表（與工作儲存同一動作產出，稽核
-  鏈無中間空缺）。詳見 `research-plan.md`「資料儲存與模型」。
-- **設定**經 pydantic-settings 統一：`.env`（gitignored，範本
-  `tools/.env.example`）供應 `SQLALCHEMY_DATABASE_URL` 與
-  `ANTHROPIC_API_KEY`，絕不寫入 repo。
-- **CLAUDE.md 極簡**：實測證實 Claude Code subagent 會繼承專案
-  CLAUDE.md 全文，故其中只放工作流程規則，領域知識一律放
-  `docs/`（subagent 不會自動讀到）。

docs/reliability.md

@@ -1,8 +1,7 @@
 # 信度:量測方式與結果
 
-（2026-08-16 量測。對象為步驟 3 的三份執行歸檔
+（2026-08-16 量測。對象為步驟 3a 與步驟 5d 各三次執行的
-`runs/03-code/run1`–`run3`,與步驟 5-4 的三份執行歸檔
+原始輸出;數字由原始輸出直接計算。）
-`runs/05-04-annotate/run1`–`run3`;數字由原始輸出直接計算。）
 
 ## 一、本研究的信度是什麼
 
@@ -34,11 +33,7 @@ Krippendorff 依產生資料的設計,把信度分成三型
 **精密度(precision)**,而**準確度(accuracy)** 未經量測。
 若儀器是確定性的,重複性無須量測;但 LLM 不是——同樣的提示、
 同樣的輸入,三次會給出不同結果,因此重複性是必須報告的儀器
-規格,而非慣例儀式。
+規格。
-
-這一點與 LLM 標註的近期文獻一致:同一模型重複取樣量測的是
-自我一致性(self-consistency),而重複執行後取多數決可提升
-標註穩定度(如 Prompt Stability Scoring,arXiv:2407.02039)。
 
 ## 二、三個常用指標
 
@@ -67,7 +62,7 @@ Krippendorff 依產生資料的設計,把信度分成三型
 
 ## 三、實測結果
 
-### 步驟 3(編碼:883 首 × 101 個碼 = 89,183 格,執行三次)
+### 步驟 3a(編碼:883 首 × 101 個碼 = 89,183 格,執行三次)
 
 | 兩次執行 | 百分比一致率 | Cohen's κ | Jaccard |
 |---|---:|---:|---:|
@@ -84,7 +79,7 @@ Krippendorff 依產生資料的設計,把信度分成三型
 (此處由原始輸出計算,未套用 936 筆人工校對表;論文引用的
 定案數 14,664 為校對後的結果,差 1 筆。)
 
-### 步驟 5-4(樣態標註:111 首 × 各自適用樣態 = 3,708 格,執行三次)
+### 步驟 5d(樣態標註:111 首 × 各自適用樣態 = 3,708 格,執行三次)
 
 | 兩次執行 | 百分比一致率 | Cohen's κ | Jaccard |
 |---|---:|---:|---:|
@@ -106,22 +101,16 @@ Jaccard(只看「有標到」的格子,忽略雙方都沒標的)約 0.89–0.91,
 比百分比一致率低而更誠實——因為 82% 的格子是雙方都判「無」,
 那些一致並不費力。
 
-## 四、隨機性的規模與三票制的作用
+## 四、隨機性的規模
 
 信度數字回答的實際問題是:**這台儀器的隨機性,大到會不會
 改變結論?**
 
-步驟 3 的 89,183 格中,搖擺的格子共 2,473 格(兩票 1,163、
+步驟 3a 的 89,183 格中,搖擺的格子共 2,473 格(兩票 1,163、
 一票 1,310),佔 2.8%。若只執行一次即定案,這批格子當中約
 一半會成為誤收、另一半會成為漏收。三票多數決把兩票以上者
 收入、一票者剔除,處理的正是這一批。
 
-三票制並非消除隨機性,而是**把隨機性往案例原本的傾向推**。
-設某個邊緣案例的符合程度為 p,單次執行以機率 p 標出,三次
-多數決則以 p³+3p²(1−p) 標出:p=0.9 者由 0.9 提高到 0.972,
-p=0.1 者由 0.1 壓低到 0.028,而 p=0.5 者仍是 0.5——真正
-模稜兩可的案例,任何票制都救不了。
-
 以此規模判斷,結論層的三條帶狀結構(女性力量與陽剛群共現、
 與脆弱群互斥、與厭女群獨立)不可能由這個量級的雜訊翻轉。
 反過來說,若一致率只有 0.6,同一組結論就不能採信——信度
@@ -146,8 +135,8 @@ p=0.1 者由 0.1 壓低到 0.028,而 p=0.5 者仍是 0.5——真正
 ## 六、計算方式與依據
 
 - **分析單位**:每一個「(歌曲, 編碼)」格為一個單位,是非題
-  (有標／無標)。步驟 3 為 883 首 × 101 碼 = 89,183 格;
+  (有標／無標)。步驟 3a 為 883 首 × 101 碼 = 89,183 格;
-  步驟 5-4 為各歌適用樣態數合計 3,708 格。
+  步驟 5d 為各歌適用樣態數合計 3,708 格。
 - **Cohen's κ**:兩次執行的 2×2 表,κ=(p_o−p_e)/(1−p_e),
   p_e 由兩次各自的邊際比例相乘求得。
 - **Krippendorff's α(名目、三位「編碼者」、無缺漏)**:

docs/research-plan.md

@@ -11,81 +11,33 @@
   **不與先導研究做比較**；
   全文數字一律以正式研究結果為準。先導研究僅作為假說的
   內部來源，記於決策日誌，不進入論文敘事。
-- **執行原則**：主會話只做討論；所有分析由 deterministic
+- **執行原則**：主會話只做討論，分析一律由確定性程序
-  script 執行。LLM 步驟以 Python script 呼叫 Anthropic
+  執行。可逐項機械比對的判斷（步驟 3 編碼、步驟 4 選群、
-  Messages API（個人 Console 帳號、Batch API 五折），定義
+  步驟 5d 樣態標註）採「三次獨立執行＋多數決」定案，
-  檔逐字作為 system prompt。可逐項機械比對的判斷（步驟 3
+  自由生成（步驟 1 自由標註）採兩次執行全數進池，步驟
-  編碼、步驟 4 選群、步驟 5-4 樣態標註）採「同一定義檔
+  5a 至 5c 定為質性閱讀協定；程序細節見 `methodology.md`，
-  獨立執行三次＋多數決」，計票由確定性子命令完成。自由
+  工作規範見 `conventions.md`。
-  生成（步驟 1 自由標註）兩次執行全數進池。步驟 5-1 至
-  5-3 為質性閱讀協定：三次獨立閱讀為分析者三角檢核、
-  逐首整合、樣態統整產出草稿，不適用投票與仲裁。詞彙表
-  不經 LLM，由詞向量嵌入＋確定性分群產生。驗證結果不符
-  預期則修訂定義檔重跑該循環，絕不手改結果。
 - **提示詞只定格式、不定語意**：研究對象是通用 LLM 以其
   網路語料知識背景所做的自然編碼與閱讀，其結果本身是
   批判對象。定義檔只規定任務形狀（輸入、數量範圍、輸出
   格式），不給任何主題的定義、判準或範例。LLM 判斷一律
   要求逐項引述歌詞原句，作為檢視偏差的依據。
-- **模型**：步驟 1、3 用 `claude-sonnet-4-6`（temperature=0、
+- **模型**：步驟 1、3 用 `claude-sonnet-4-6`；步驟 4、5
-  thinking 關閉）；步驟 4、5 用 `claude-fable-5`（兩參數
+  用 `claude-fable-5`——實測發現 sonnet 將「Women Power」
-  不適用，均不送出；取樣變異由多數決或整合吸收）——實測
+  拆讀為 women＋power 的組合語意，fable-5 讀為詞彙化
-  發現 sonnet 將「Women Power」拆讀為 women＋power 的組合
+  概念，語意層任務因此換用 fable-5（經過見決策日誌）。
-  語意，fable-5 讀為詞彙化概念，語意層任務因此換用
-  fable-5（經過見決策日誌）。
 - **信度與效度**：三次獨立執行量測穩定性（intra-rater
   reliability，可報告兩兩一致率）；封閉母體全量檢查取代
   抽樣防衛。可重現性定義為「程序透明＋可稽核」：公開定義
   檔、記錄 model ID 與執行時間、保存全部原始輸出。
 
-## 資料儲存與模型
+## 範圍性定案
 
-- **Commit 判準**：凡能由「committed 的輸入＋committed 的
-  程式」決定性再生者，不 commit；凡不能者——源頭資料、
-  外部世界的捕捉（Wikidata 快照、LLM 原始輸出）、人工
-  著作——一律以文字格式 commit。格式跟著層次走。
-- **分層**：
-  - 源頭：原始榜單 CSV（進 git）。
-  - 捕捉：Wikidata 快照 CSV、`runs/` JSONL（皆進 git）；
-    歌詞 `.txt` 快取（版權因素 gitignored，為已知的稽核
-    缺口）。
-  - 人工：`data/manual/`（僅研究者親手寫入：編碼修正、
-    演唱者性別修正）。
-  - 工作儲存：SQLite 單檔（`tools/instance/`，generated、
-    不進 git），SQLAlchemy 2.0 typed ORM 定義 schema，
-    設定經 pydantic-settings（`.env` 供應
-    `SQLALCHEMY_DATABASE_URL` 與 `ANTHROPIC_API_KEY`）。
-    歌詞全文入 DB（不進 git 故無版權疑慮）。
-  - 衍生：`build-db` 建置工作儲存的同一動作產出人讀報表
-    （`data/derived/`，進 git），與 SQLite 同交易語意。
-  - 報表：論文引用的定案表進 `results/`——
-    `codings.csv`（步驟 3 定案編碼）、`groups.csv`
-    （步驟 4 定案編碼群）、`patterns.csv`（步驟 5-3 定案
-    樣態表）、`annotations.csv`（步驟 5-4 定案歌×樣態
-    矩陣）。衍生與報表為「可再生仍 commit」的例外，理由：
-    引用穩定性、審稿人零門檻、撰稿期數字變動可 diff。
-- **資料模型**：`songs`（含 lyrics、`performer_gender`——
-  演唱聲音之性別，由署名藝人之 Wikidata 性別推導後套用
-  `data/manual/performer-gender-corrections.csv` 手工修正）、
-  `chart_entries`、`artists`、`song_artists`、`codings`、
-  `groups`（語意編碼群）、`patterns`（深讀樣態）、
-  `annotations`（歌×樣態定案矩陣）。領域不變量（恰 1000
-  筆榜單、每歌至少一 primary 歌手等）檢查內建於
-  `build-db`，違規即建置失敗。
 - **歌手背景防火牆**：歌手背景資料只進人工解讀階段，
   **絕不進 LLM 輸入**——LLM 任務的 user message 維持
   歌詞-only 或管線中間產物-only，避免光環偏誤。
-- **Pilot 歌詞沿用（私人匯入，不進發布管線）**：先導研究
+- **Pilot 歌詞沿用**：先導歌詞沿用之來歷與細節詳見該檔
-  捕捉檔（lyrics.json，684 首，2018–2025）以 `excludes/`
+  （`pilot-study.md`）。
-  的私人腳本匯入歌詞快取；讀者的重現路徑純粹是
-  `fetch-lyrics`；沿用之事實記於
-  `data/captures/lyrics-provenance.csv`（進 git）。
-- **子命令**（`pop-fem-audit-tools <cmd>`）：`build-db`
-  （`--codings`、`--groups`、`--gender-corrections`、
-  `--patterns`、`--annotations`）、`cluster-keywords`、
-  `export-llm-input`、`fetch-artists`、`fetch-lyrics`、
-  `run-llm`（`--model` 於模型登錄表中擇一）、
-  `tally-codings`、`tally-groups`、`tally-annotations`。
 
 ## 分析管線（五步驟，全部完成）
 
@@ -96,26 +48,19 @@
    階層式聚合分群 k=100，組名取 medoid；再併入研究者先驗
    主題詞 `women-power`（據實揭露的儀器介入），共 101 碼。
 3. **編碼（步驟 3）**：以定稿詞彙表對全 883 首編碼，逐
-   標籤附引述，×3＋多數決（`tally-codings`），定案
+   標籤附引述，×3＋多數決，產出定案編碼表。「女性力量」
-   `results/codings.csv`。「女性力量」候選集（wp 66 首、
+   候選集（wp 66 首、fe 144 首、wp∪fe 145 首）由此浮現。
-   fe 144 首、wp∪fe 145 首）由此浮現。
 4. **語意編碼群（步驟 4）**：LLM 依編碼字面語意將 101 碼
    選入研究者指定的四個主題群（women-power／misogyny／
-   masculine／vulnerable），×3＋多數決（`tally-groups`），
+   masculine／vulnerable），×3＋多數決，產出定案分群表；
-   定案 `results/groups.csv`；wp/fe 與各編碼、各編碼群之
+   wp/fe 與各編碼、各編碼群之關聯統計（BH-FDR 校正）入
-   關聯統計（BH-FDR 校正）入論文。
+   論文。
 5. **女性主義問題之質性深讀（步驟 5）**：對 wp∪fe 145 首
-   ——5-1 逐首盲讀（僅歌詞全文）×3；5-2 逐首整合（收斂
+   ——5a 逐首盲讀（僅歌詞全文）×3；5b 逐首整合（收斂
-   註記、主清單限兩讀以上）；5-3 樣態統整（全體基底＋
+   註記、主清單限兩讀以上）；5c 樣態統整（全體基底＋
-   男聲／女聲／混合三個發話脈絡分組）；5-4 樣態標註——
+   男聲／女聲／混合三個發話脈絡分組）；5d 樣態標註——
-   以分組樣態表逐首標註「有問題」的 111 首，×3＋多數決
+   以分組樣態表逐首標註「有問題」的 111 首，×3＋多數決，
-   （`tally-annotations`），定案 `results/patterns.csv` 與
+   產出定案樣態表與歌×樣態定案矩陣。
-   `results/annotations.csv`。
-
-定義檔命名 `prompts/<步>-<次步>-<task>.md`，編號的所指是
-工序而非定義檔（確定性的第 2 步無定義檔仍佔編號）；檔名
-不帶版本號——版本即 git 歷史；每次執行的定義檔快照隨
-`runs/` 自我完備，token 費用逐筆記於 `docs/run-costs.md`。
 
 ## 時程與剩餘工作
 
@@ -124,12 +69,3 @@
   污染」結論（步驟 5 素材已備）、信度說明、fe 與新自由
   主義敘事之詮釋標註。
 
-## 其他已定案事項
-
-- 歌詞受版權保護：完整歌詞不進 git
-  （`data/captures/lyrics/` gitignored），論文與 repo 只留
-  分析所引摘錄。
-- 工程支援（寫 script、style check、審稿）用公司訂閱的
-  Claude Code；進論文的分析 token 由個人 API 帳號支付。
-- Positionality statement 寫入論文（53 歲、長年婦運者、
-  資深工程師、離開學術圈多年、因整理榜單而起）。

docs/run-costs.md

@@ -10,57 +10,61 @@ $3／$15、opus-4-6 $5／$25、opus-5 與 fable-5 $10／$50）
 被 `--replace` 取代的執行以「已取代」標記，數字保留供
 總支出核算。
 
-| 日期 | 步驟 | 執行 | 模型 | 批次 ID | 耗時 | input | output | 費用 (USD) | 狀態 |
+「步驟」欄的名稱於 2026-08-18 一律換為現行編號（見
-|---|---|---|---|---|---|---:|---:|---:|---|
+`decision-log.md` 當日條目），執行當時的名稱記於末欄
-| 2026-08-04 | 01-01-01-tag | run1 | claude-sonnet-4-6 | msgbatch_01PTACDQMr8M6ahnshbedjtB | 6 分 27 秒 | 763,318 | 338,661 | $3.68 | 已取代（浮水印清洗與防圍欄修訂後重跑） |
+「原步驟名」，未曾改名者留空。已刪除的工序——步驟 2 的
-| 2026-08-05 | 01-01-01-tag | run1 | claude-sonnet-4-6 | msgbatch_01TikJNd2pZVxzQ8SaybVthu | 4 分 57 秒 | 774,604 | 323,651 | $3.59 | 已取代（合法 JSON 修訂後重跑） |
+LLM 合併嘗試（`01-02-01-merge`，詞彙表改採詞向量分群時
-| 2026-08-05 | 01-01-01-tag | run1 | claude-sonnet-4-6 | msgbatch_01JFBCNqnu1cwXyqmEKLHQYF | 4 分 30 秒 | 790,370 | 328,227 | $3.65 | 已取代（song-288 平台失敗，整批重跑驗證） |
+放棄）與步驟 3 的仲裁（`03-02-arbitration`，改採三票
-| 2026-08-05 | 01-01-01-tag | run1 | claude-sonnet-4-6 | msgbatch_01VSDneWuSbf8mShA32jiWrX | 6 分 15 秒 | 790,370 | 326,193 | $3.63 | 已取代（措辭修訂後全體重跑） |
+多數決時刪除）——不另發新名，步驟欄標「已廢棄」，僅存
-| 2026-08-05 | 01-01-01-tag | run1-rescue-288 | claude-sonnet-4-6 | msgbatch_019Hq6bNVXVmp4DjRZ2cgVda | 2 分 5 秒 | 1,015 | 376 | $0.01 | 已取代（措辭修訂後全體重跑，該首原生通過）|
+本表供總支出核算。
-| 2026-08-05 | 01-01-01-tag | run1 | claude-sonnet-4-6 | msgbatch_01VgZ77KAPGWmuu3PnQFqZ7Q | 7 分 1 秒 | 794,913 | 326,435 | $3.64 | 現行 |
-| 2026-08-05 | 01-01-01-tag | run2 | claude-sonnet-4-6 | msgbatch_01TLFey3L4fimKxcebTZQGYn | 4 分 21 秒 | 794,913 | 328,324 | $3.65 | 現行 |
-| 2026-08-05 | 01-02-01-merge | run1 | claude-sonnet-4-6 | msgbatch_01BvYMFmH8zrUq9SNxWSNba7 | 9 分 14 秒 | 46,454 | 60,974 | $0.53 | 已取代（完整分割驗證不過：漏 366、重複分派 511、撞名 4） |
-| 2026-08-05 | 01-02-01-merge | run1 | claude-opus-5 | msgbatch_018TkYpdT3DsZqq2q94FQCUN | — | 0 | 0 | $0.00 | 拒收（temperature 已棄用） |
-| 2026-08-05 | 01-02-01-merge | run1 | claude-opus-4-8 | msgbatch_018Wuux4adz5mWSvzJjfgLxb | — | 0 | 0 | $0.00 | 拒收（temperature 已棄用） |
-| 2026-08-05 | 01-02-01-merge | run1 | claude-opus-4-6 | msgbatch_01Lmt2j1Txyof9CctUK7CbHA | 12 分 19 秒 | 46,454 | 54,864 | $0.80 | 已取代（圍欄違規；漏 190、發明 151、重複 11） |
-| 2026-08-05 | 01-02-01-merge | run1 | claude-opus-5 | msgbatch_012SEfebTnjU5uWN6hhfnhhh | 8 分 45 秒 | 63,368 | 64,000 | $1.92 | 已取代（64k 截斷） |
-| 2026-08-05 | 01-02-01-merge | run1 | claude-opus-5 | msgbatch_01Wi9xEoQKCy1nZK6myVNg77 | 11 分 34 秒 | 63,368 | 70,175 | $2.07 | 已取代（驗證不過：漏 59、發明 140、重複 3） |
-| 2026-08-05 | 01-02-01-merge | run1 | claude-fable-5 | msgbatch_01M15hUs8P9FDTqWvAHdc1cA | 32 分 45 秒 | 63,368 | 107,721 | $3.01 | 現行歸檔（JSON 語法毀損，驗證不過） |
-| 2026-08-05 | 01-02-01-merge | run1 | claude-fable-5 | msgbatch_012stRP28uuQDQwobm41HMLh | — | 0 | 0 | $0.00 | 拒收（thinking.type.enabled 不支援） |
-| 2026-08-05 | 01-02-01-merge | run1 | claude-fable-5 | msgbatch_01KpS7AMx1zAHTcNujgGsouJ | 26 分 28 秒 | 63,368 | 128,000 | $3.52 | 現行歸檔（effort max：128k 全耗於推理，正文空白） |
-| 2026-08-05 | 03-01-code | run1 | claude-sonnet-4-6 | msgbatch_01L7VepTxSNz5vruvzzjuL2c | 6 分 36 秒 | 1,247,264 | 492,730 | $5.57 | 36 首輸出遭內容過濾攔阻，待修訂重跑 |
-| 2026-08-05 | 03-01-code | 引述減量實驗（36 首）| claude-sonnet-4-6 | msgbatch_01Hskdhi2DkudYmgZhhgFts7 | 3 分 51 秒 | 56,458 | 11,074 | $0.17 | 實驗：每碼一行引述，36 首全數通過過濾；歸檔不入 repo |
-| 2026-08-05 | 03-01-code | run1 | claude-sonnet-4-6 | msgbatch_01GG5Ez9KT1pPwtQaY4sW7tv | 4 分 35 秒 | 1,293,724 | 269,166 | $3.96 | 過濾零攔阻；song-168、song-590 因餘額用盡失敗，song-775 拒答 |
-| 2026-08-05 | 03-01-code | 101 碼樹狀探測（148 首）| claude-sonnet-4-6 | msgbatch_017jo3E5gWVks9b39iWMTqz3 | 2 小時 11 分 | 385,270 | 79,690 | $0.87 | 實驗：k=100 葉碼＋women-power；零違規碼； 歸檔不入 repo |
-| 2026-08-06 | 命名實驗（100 組）| — | claude-sonnet-4-6 | msgbatch_01XSi1YtWzdVyYUzYh7DQRWg | 3 分 5 秒 | 55,090 | 1,168 | $0.09 | 實驗：LLM 命名對照 medoid，未採用；歸檔不入 repo |
-| 2026-08-06 | 命名實驗（100 組）| — | claude-fable-5 | msgbatch_01Y8SJj1h1ZuSQRErhqkvgZE | 3 分 2 秒 | 76,211 | 2,049 | $0.43 | 實驗：同上，加禁用 themes；未採用；歸檔不入 repo |
-| 2026-08-06 | 03-01-code | run1 | claude-sonnet-4-6 | msgbatch_01GS1opvurvsf62oknnQxhtx | 3 分 46 秒 | 1,625,458 | 362,759 | $5.16 | 現行（101 碼；883 首全數有效，零攔阻） |
-| 2026-08-06 | 03-01-code | run2 | claude-sonnet-4-6 | msgbatch_01CxnwNLWzZbRpZK7UAdpb8i | 5 分 22 秒 | 1,625,458 | 364,098 | $5.17 | 現行（101 碼；883 首全數有效，零攔阻） |
-| 2026-08-06 | 03-02-arbitration | — | claude-sonnet-4-6 | msgbatch_019xcQXwrbwDGFc9nE5M8AjE | 4 分 0 秒 | 763,193 | 42,389 | $1.46 | 已取代（2 首遭內容過濾攔阻、13 首輸出夾帶散文；定義檔修訂後重跑） |
-| 2026-08-06 | 03-02-arbitration | — | claude-sonnet-4-6 | msgbatch_01N7bDbXRSfAVUzzaj2thKeR | 4 分 20 秒 | 781,037 | 39,315 | $1.47 | 現行（644 首全數有效，零攔阻；保留 1,481／送裁 1,699） |
-| 2026-08-06 | 03-code | run3 | claude-sonnet-4-6 | msgbatch_01KnkCaGETnFJrPddrxZTYHA | 6 分 26 秒 | 1,625,458 | 363,840 | $5.17 | 現行（101 碼；883 首全數有效，零攔阻） |
-| 2026-08-14 | 04-group | run1 | claude-sonnet-4-6 | msgbatch_01UPNedog6feQzJ9WVfSAxBD | 1 分 1 秒 | 3,129 | 470 | $0.01 | 已取代（僅 3 群；改納 women-power 群後重跑） |
-| 2026-08-14 | 04-group | run1 | claude-sonnet-4-6 | msgbatch_01KntJgxdicStaMjNL12P3zi | 1 分 25 秒 | 4,170 | 558 | $0.01 | 已取代（改以 claude-fable-5 執行；vulnerable 輸出含詞彙表外碼 1 筆；歸檔備份於 excludes/experiments/） |
-| 2026-08-14 | 04-group | run1 | claude-fable-5 | msgbatch_01Mr6goBb2Efa4YrqCprbP4U | 55 秒 | 5,553 | 2,049 | $0.08 | 現行（4 群；零違規碼；temperature 與 thinking 參數不適用於本模型，未送出） |
-| 2026-08-14 | 04-group | run2 | claude-fable-5 | msgbatch_01C17xW3YBefThTYZ83g7KkL | 1 分 21 秒 | 5,553 | 1,891 | $0.08 | 現行（4 群；零違規碼） |
-| 2026-08-14 | 04-group | run3 | claude-fable-5 | msgbatch_01DveEMYyjCAYe6wxpcCD87V | 2 分 9 秒 | 5,553 | 1,975 | $0.08 | 現行（4 群；零違規碼） |
-| 2026-08-15 | 05-01-read | run1 | claude-fable-5 | msgbatch_018NJejke75o7kzKyjZeeKwi | 2 分 12 秒 | 189,607 | 181,299 | $5.48 | 現行（144/145 有效；song-444 平台錯誤，單筆補送） |
-| 2026-08-15 | 05-01-read | run1-rescue-444 | claude-fable-5 | msgbatch_01DgygMm6iYn2qEnyqHkTMfa | 1 分 52 秒 | 1,852 | 1,557 | $0.05 | 現行（run1 之 song-444 單筆補送，成功） |
-| 2026-08-15 | 05-01-read | run2 | claude-fable-5 | msgbatch_011txWpWDtdkNXNiPA74ZUfQ | 15 分 25 秒 | 191,459 | 185,630 | $5.60 | 現行（145/145 有效） |
-| 2026-08-15 | 05-01-read | run3 | claude-fable-5 | msgbatch_01Unn7C7WCnuVoLraJRXksoG | 7 分 59 秒 | 191,459 | 190,433 | $5.72 | 現行（145/145 有效） |
-| 2026-08-15 | 05-02-consolidate | run1 | claude-fable-5 | msgbatch_01WPyH4pCCiieHs4yU3xtHzN | 2 分 28 秒 | 326,287 | 242,785 | $7.70 | 現行（145/145 有效） |
-| 2026-08-15 | 05-03-synthesize | run1 | claude-fable-5 | msgbatch_01XbNQb1SFrUxhk2ZWF7epPA | 3 分 0 秒 | 103,888 | 9,982 | $0.77 | 現行（15 個樣態；研究者審定用草稿） |
 
-| 2026-08-15 | 05-03-synthesize | run2 | claude-fable-5 | msgbatch_01FKqq8U6m5HLLsx7KpQES26 | 3 分 4 秒 | 13,573 | 6,883 | $0.24 | 現行（男聲群 14 首；13 個樣態） |
+| 日期 | 步驟 | 執行 | 模型 | 批次 ID | 耗時 | input | output | 費用 (USD) | 狀態 | 原步驟名 |
-| 2026-08-15 | 05-03-synthesize | run3 | claude-fable-5 | msgbatch_017tfhiRTnaxqCp6YFyoivsu | 4 分 4 秒 | 60,727 | 8,519 | $0.52 | 現行（女聲群 98 首；14 個樣態） |
+|---|---|---|---|---|---|---:|---:|---:|---|---|
-| 2026-08-15 | 05-03-synthesize | run4 | claude-fable-5 | msgbatch_01SNjTLiNPEba4LW2tcUAcTJ | 4 分 3 秒 | 29,958 | 9,793 | $0.39 | 現行（混合群 30 首；16 個樣態） |
+| 2026-08-04 | 1-tag | run1 | claude-sonnet-4-6 | msgbatch_01PTACDQMr8M6ahnshbedjtB | 6 分 27 秒 | 763,318 | 338,661 | $3.68 | 已取代（浮水印清洗與防圍欄修訂後重跑） | 01-01-01-tag |
-
+| 2026-08-05 | 1-tag | run1 | claude-sonnet-4-6 | msgbatch_01TikJNd2pZVxzQ8SaybVthu | 4 分 57 秒 | 774,604 | 323,651 | $3.59 | 已取代（合法 JSON 修訂後重跑） | 01-01-01-tag |
-| 2026-08-15 | 05-04-annotate | run1 | claude-fable-5 | msgbatch_01H9gXXhNbhZG6gzZtvMHxhh | 3 分 9 秒 | 815,492 | 300,901 | $11.60 | 現行（109/111 有效；song-177、song-199 遭 max_tokens 截斷，單筆補送） |
+| 2026-08-05 | 1-tag | run1 | claude-sonnet-4-6 | msgbatch_01JFBCNqnu1cwXyqmEKLHQYF | 4 分 30 秒 | 790,370 | 328,227 | $3.65 | 已取代（song-288 平台失敗，整批重跑驗證） | 01-01-01-tag |
-| 2026-08-15 | 05-04-annotate | run1-rescue-177-199 | claude-fable-5 | msgbatch_01YEdroZ51b8btPfG1aqPFCD | 3 分 5 秒 | 17,802 | 11,601 | $0.38 | 現行（run1 之 2 筆補送，max_tokens 提為 16000，成功） |
+| 2026-08-05 | 1-tag | run1 | claude-sonnet-4-6 | msgbatch_01VSDneWuSbf8mShA32jiWrX | 6 分 15 秒 | 790,370 | 326,193 | $3.63 | 已取代（措辭修訂後全體重跑） | 01-01-01-tag |
-| 2026-08-15 | 05-04-annotate | — | claude-fable-5 | msgbatch_01G67STqZsHpnhedat2HRU85 | 3 分 0 秒 | 17,802 | 14,162 | $0.44 | 誤送（補送批次輪詢中斷後誤判死亡而重送；原批次自行完成並歸檔為現行，本批次取消不及、兩筆皆完成，結果棄用） |
+| 2026-08-05 | 1-tag | run1-rescue-288 | claude-sonnet-4-6 | msgbatch_019Hq6bNVXVmp4DjRZ2cgVda | 2 分 5 秒 | 1,015 | 376 | $0.01 | 已取代（措辭修訂後全體重跑，該首原生通過）| 01-01-01-tag |
-
+| 2026-08-05 | 1-tag | run1 | claude-sonnet-4-6 | msgbatch_01VgZ77KAPGWmuu3PnQFqZ7Q | 7 分 1 秒 | 794,913 | 326,435 | $3.64 | 現行 | 01-01-01-tag |
-| 2026-08-15 | 05-04-annotate | run2 | claude-fable-5 | msgbatch_01VVJpsfidNtF66GnHQdWnGR | 9 分 16 秒 | 815,492 | 306,582 | $11.74 | 現行（111/111 有效；max_tokens 16000，零截斷） |
+| 2026-08-05 | 1-tag | run2 | claude-sonnet-4-6 | msgbatch_01TLFey3L4fimKxcebTZQGYn | 4 分 21 秒 | 794,913 | 328,324 | $3.65 | 現行 | 01-01-01-tag |
-
+| 2026-08-05 | 已廢棄 | run1 | claude-sonnet-4-6 | msgbatch_01BvYMFmH8zrUq9SNxWSNba7 | 9 分 14 秒 | 46,454 | 60,974 | $0.53 | 已取代（完整分割驗證不過：漏 366、重複分派 511、撞名 4） | 01-02-01-merge |
-| 2026-08-15 | 05-04-annotate | run3 | claude-fable-5 | msgbatch_01AsYvZd8rsTYVfUYLzWfeMK | 4 分 12 秒 | 815,492 | 298,512 | $11.54 | 現行（111/111 有效；max_tokens 16000，零截斷） |
+| 2026-08-05 | 已廢棄 | run1 | claude-opus-5 | msgbatch_018TkYpdT3DsZqq2q94FQCUN | — | 0 | 0 | $0.00 | 拒收（temperature 已棄用） | 01-02-01-merge |
+| 2026-08-05 | 已廢棄 | run1 | claude-opus-4-8 | msgbatch_018Wuux4adz5mWSvzJjfgLxb | — | 0 | 0 | $0.00 | 拒收（temperature 已棄用） | 01-02-01-merge |
+| 2026-08-05 | 已廢棄 | run1 | claude-opus-4-6 | msgbatch_01Lmt2j1Txyof9CctUK7CbHA | 12 分 19 秒 | 46,454 | 54,864 | $0.80 | 已取代（圍欄違規；漏 190、發明 151、重複 11） | 01-02-01-merge |
+| 2026-08-05 | 已廢棄 | run1 | claude-opus-5 | msgbatch_012SEfebTnjU5uWN6hhfnhhh | 8 分 45 秒 | 63,368 | 64,000 | $1.92 | 已取代（64k 截斷） | 01-02-01-merge |
+| 2026-08-05 | 已廢棄 | run1 | claude-opus-5 | msgbatch_01Wi9xEoQKCy1nZK6myVNg77 | 11 分 34 秒 | 63,368 | 70,175 | $2.07 | 已取代（驗證不過：漏 59、發明 140、重複 3） | 01-02-01-merge |
+| 2026-08-05 | 已廢棄 | run1 | claude-fable-5 | msgbatch_01M15hUs8P9FDTqWvAHdc1cA | 32 分 45 秒 | 63,368 | 107,721 | $3.01 | 現行歸檔（JSON 語法毀損，驗證不過） | 01-02-01-merge |
+| 2026-08-05 | 已廢棄 | run1 | claude-fable-5 | msgbatch_012stRP28uuQDQwobm41HMLh | — | 0 | 0 | $0.00 | 拒收（thinking.type.enabled 不支援） | 01-02-01-merge |
+| 2026-08-05 | 已廢棄 | run1 | claude-fable-5 | msgbatch_01KpS7AMx1zAHTcNujgGsouJ | 26 分 28 秒 | 63,368 | 128,000 | $3.52 | 現行歸檔（effort max：128k 全耗於推理，正文空白） | 01-02-01-merge |
+| 2026-08-05 | 3a-code | run1 | claude-sonnet-4-6 | msgbatch_01L7VepTxSNz5vruvzzjuL2c | 6 分 36 秒 | 1,247,264 | 492,730 | $5.57 | 36 首輸出遭內容過濾攔阻，待修訂重跑 | 03-01-code |
+| 2026-08-05 | 3a-code | 引述減量實驗（36 首）| claude-sonnet-4-6 | msgbatch_01Hskdhi2DkudYmgZhhgFts7 | 3 分 51 秒 | 56,458 | 11,074 | $0.17 | 實驗：每碼一行引述，36 首全數通過過濾；歸檔不入 repo | 03-01-code |
+| 2026-08-05 | 3a-code | run1 | claude-sonnet-4-6 | msgbatch_01GG5Ez9KT1pPwtQaY4sW7tv | 4 分 35 秒 | 1,293,724 | 269,166 | $3.96 | 過濾零攔阻；song-168、song-590 因餘額用盡失敗，song-775 拒答 | 03-01-code |
+| 2026-08-05 | 3a-code | 101 碼樹狀探測（148 首）| claude-sonnet-4-6 | msgbatch_017jo3E5gWVks9b39iWMTqz3 | 2 小時 11 分 | 385,270 | 79,690 | $0.87 | 實驗：k=100 葉碼＋women-power；零違規碼； 歸檔不入 repo | 03-01-code |
+| 2026-08-06 | 命名實驗（100 組）| — | claude-sonnet-4-6 | msgbatch_01XSi1YtWzdVyYUzYh7DQRWg | 3 分 5 秒 | 55,090 | 1,168 | $0.09 | 實驗：LLM 命名對照 medoid，未採用；歸檔不入 repo |  |
+| 2026-08-06 | 命名實驗（100 組）| — | claude-fable-5 | msgbatch_01Y8SJj1h1ZuSQRErhqkvgZE | 3 分 2 秒 | 76,211 | 2,049 | $0.43 | 實驗：同上，加禁用 themes；未採用；歸檔不入 repo |  |
+| 2026-08-06 | 3a-code | run1 | claude-sonnet-4-6 | msgbatch_01GS1opvurvsf62oknnQxhtx | 3 分 46 秒 | 1,625,458 | 362,759 | $5.16 | 現行（101 碼；883 首全數有效，零攔阻） | 03-01-code |
+| 2026-08-06 | 3a-code | run2 | claude-sonnet-4-6 | msgbatch_01CxnwNLWzZbRpZK7UAdpb8i | 5 分 22 秒 | 1,625,458 | 364,098 | $5.17 | 現行（101 碼；883 首全數有效，零攔阻） | 03-01-code |
+| 2026-08-06 | 已廢棄 | — | claude-sonnet-4-6 | msgbatch_019xcQXwrbwDGFc9nE5M8AjE | 4 分 0 秒 | 763,193 | 42,389 | $1.46 | 已取代（2 首遭內容過濾攔阻、13 首輸出夾帶散文；定義檔修訂後重跑） | 03-02-arbitration |
+| 2026-08-06 | 已廢棄 | — | claude-sonnet-4-6 | msgbatch_01N7bDbXRSfAVUzzaj2thKeR | 4 分 20 秒 | 781,037 | 39,315 | $1.47 | 現行（644 首全數有效，零攔阻；保留 1,481／送裁 1,699） | 03-02-arbitration |
+| 2026-08-06 | 3a-code | run3 | claude-sonnet-4-6 | msgbatch_01KnkCaGETnFJrPddrxZTYHA | 6 分 26 秒 | 1,625,458 | 363,840 | $5.17 | 現行（101 碼；883 首全數有效，零攔阻） | 03-code |
+| 2026-08-14 | 4-group | run1 | claude-sonnet-4-6 | msgbatch_01UPNedog6feQzJ9WVfSAxBD | 1 分 1 秒 | 3,129 | 470 | $0.01 | 已取代（僅 3 群；改納 women-power 群後重跑） | 04-group |
+| 2026-08-14 | 4-group | run1 | claude-sonnet-4-6 | msgbatch_01KntJgxdicStaMjNL12P3zi | 1 分 25 秒 | 4,170 | 558 | $0.01 | 已取代（改以 claude-fable-5 執行；vulnerable 輸出含詞彙表外碼 1 筆；歸檔另行私人備份，不入版本庫） | 04-group |
+| 2026-08-14 | 4-group | run1 | claude-fable-5 | msgbatch_01Mr6goBb2Efa4YrqCprbP4U | 55 秒 | 5,553 | 2,049 | $0.08 | 現行（4 群；零違規碼；temperature 與 thinking 參數不適用於本模型，未送出） | 04-group |
+| 2026-08-14 | 4-group | run2 | claude-fable-5 | msgbatch_01C17xW3YBefThTYZ83g7KkL | 1 分 21 秒 | 5,553 | 1,891 | $0.08 | 現行（4 群；零違規碼） | 04-group |
+| 2026-08-14 | 4-group | run3 | claude-fable-5 | msgbatch_01DveEMYyjCAYe6wxpcCD87V | 2 分 9 秒 | 5,553 | 1,975 | $0.08 | 現行（4 群；零違規碼） | 04-group |
+| 2026-08-15 | 5a-read | run1 | claude-fable-5 | msgbatch_018NJejke75o7kzKyjZeeKwi | 2 分 12 秒 | 189,607 | 181,299 | $5.48 | 現行（144/145 有效；song-444 平台錯誤，單筆補送） | 05-01-read |
+| 2026-08-15 | 5a-read | run1-rescue-444 | claude-fable-5 | msgbatch_01DgygMm6iYn2qEnyqHkTMfa | 1 分 52 秒 | 1,852 | 1,557 | $0.05 | 現行（run1 之 song-444 單筆補送，成功） | 05-01-read |
+| 2026-08-15 | 5a-read | run2 | claude-fable-5 | msgbatch_011txWpWDtdkNXNiPA74ZUfQ | 15 分 25 秒 | 191,459 | 185,630 | $5.60 | 現行（145/145 有效） | 05-01-read |
+| 2026-08-15 | 5a-read | run3 | claude-fable-5 | msgbatch_01Unn7C7WCnuVoLraJRXksoG | 7 分 59 秒 | 191,459 | 190,433 | $5.72 | 現行（145/145 有效） | 05-01-read |
+| 2026-08-15 | 5b-consolidate | run1 | claude-fable-5 | msgbatch_01WPyH4pCCiieHs4yU3xtHzN | 2 分 28 秒 | 326,287 | 242,785 | $7.70 | 現行（145/145 有效） | 05-02-consolidate |
+| 2026-08-15 | 5c-synthesize | run1 | claude-fable-5 | msgbatch_01XbNQb1SFrUxhk2ZWF7epPA | 3 分 0 秒 | 103,888 | 9,982 | $0.77 | 現行（15 個樣態；研究者審定用草稿） | 05-03-synthesize |
+| 2026-08-15 | 5c-synthesize | run2 | claude-fable-5 | msgbatch_01FKqq8U6m5HLLsx7KpQES26 | 3 分 4 秒 | 13,573 | 6,883 | $0.24 | 現行（男聲群 14 首；13 個樣態） | 05-03-synthesize |
+| 2026-08-15 | 5c-synthesize | run3 | claude-fable-5 | msgbatch_017tfhiRTnaxqCp6YFyoivsu | 4 分 4 秒 | 60,727 | 8,519 | $0.52 | 現行（女聲群 98 首；14 個樣態） | 05-03-synthesize |
+| 2026-08-15 | 5c-synthesize | run4 | claude-fable-5 | msgbatch_01SNjTLiNPEba4LW2tcUAcTJ | 4 分 3 秒 | 29,958 | 9,793 | $0.39 | 現行（混合群 30 首；16 個樣態） | 05-03-synthesize |
+| 2026-08-15 | 5d-annotate | run1 | claude-fable-5 | msgbatch_01H9gXXhNbhZG6gzZtvMHxhh | 3 分 9 秒 | 815,492 | 300,901 | $11.60 | 現行（109/111 有效；song-177、song-199 遭 max_tokens 截斷，單筆補送） | 05-04-annotate |
+| 2026-08-15 | 5d-annotate | run1-rescue-177-199 | claude-fable-5 | msgbatch_01YEdroZ51b8btPfG1aqPFCD | 3 分 5 秒 | 17,802 | 11,601 | $0.38 | 現行（run1 之 2 筆補送，max_tokens 提為 16000，成功） | 05-04-annotate |
+| 2026-08-15 | 5d-annotate | — | claude-fable-5 | msgbatch_01G67STqZsHpnhedat2HRU85 | 3 分 0 秒 | 17,802 | 14,162 | $0.44 | 誤送（補送批次輪詢中斷後誤判死亡而重送；原批次自行完成並歸檔為現行，本批次取消不及、兩筆皆完成，結果棄用） | 05-04-annotate |
+| 2026-08-15 | 5d-annotate | run2 | claude-fable-5 | msgbatch_01VVJpsfidNtF66GnHQdWnGR | 9 分 16 秒 | 815,492 | 306,582 | $11.74 | 現行（111/111 有效；max_tokens 16000，零截斷） | 05-04-annotate |
+| 2026-08-15 | 5d-annotate | run3 | claude-fable-5 | msgbatch_01AsYvZd8rsTYVfUYLzWfeMK | 4 分 12 秒 | 815,492 | 298,512 | $11.54 | 現行（111/111 有效；max_tokens 16000，零截斷） | 05-04-annotate |
 
 累計支出：$125.65。

tools/docs/source/changelog.rst

@@ -0,0 +1,10 @@
+Change Log
+==========
+
+
+version 1.0.0
+-------------
+
+Released 2026/8/19
+
+Initial release.

tools/docs/source/index.rst

@@ -13,6 +13,8 @@ This is a collection of supporting tools for the conference paper "流行音樂
    :maxdepth: 2
    :caption: Contents:
 
+   changelog
+
 
 Indices and tables
 ==================

tools/src/pop_fem_audit_tools/__init__.py

@@ -6,5 +6,5 @@
 """Tools for A Feminist Audit of Pop Music."""
 
 
-VERSION: str = "0.0.0"
+VERSION: str = "1.0.0"
 """The package version."""

tools/src/pop_fem_audit_tools/commands/cluster_keywords.py

@@ -24,8 +24,7 @@ keyword set for ``export-llm-input --extras`` is written as a JSON
 file holding the group name keywords plus every extra a-priori
 keyword the caller gives with the repeatable ``--extra-keyword``
 command-line option, as
-:attr:`KeywordsToMerge.KEYWORDS_TO_MERGE_JSON`; with no
+:attr:`KeywordsToMerge.KEYWORDS_TO_MERGE_JSON`.  No default
-``--extra-keyword``, it holds the group names alone.  No default
 extra keyword is ever injected; the caller supplies each one
 consciously.  Finally, the command-line choices and the
 environment that produced the numbers -- neither recoverable from
@@ -160,15 +159,8 @@ class KeywordPooler:
             if line.strip() == "":
                 continue
             record: Any = json.loads(line)
-            if not isinstance(record, dict) or "id" not in record:
-                raise ValueError(
-                    f"{path}: record without \"id\": {line}")
             if "error" in record:
                 continue
-            if "text" not in record:
-                raise ValueError(
-                    f"{path}: id {record['id']}: record without"
-                    " \"text\" or \"error\"")
             song_id: int = cls.__parse_song_id(record["id"], path)
             try:
                 keywords: Any = json.loads(
@@ -297,7 +289,8 @@ class KeywordGroups:
 class KeywordClusterer:
     """The clusterer of the pooled keywords into coding groups."""
 
-    DEFAULT_MODEL: str = "sentence-transformers/all-mpnet-base-v2"
+    DEFAULT_MODEL: ClassVar[str] \
+        = "sentence-transformers/all-mpnet-base-v2"
     """The sentence embedding model used when the caller names
     none."""
 
@@ -668,11 +661,7 @@ def parse_args(argv: list[str] | None) -> argparse.Namespace:
     parser.add_argument(
         "output_dir", type=Path,
         help="the output directory, created if missing, that"
-             f" receives {PooledKeywords.SOURCE_KEYWORDS_TXT},"
+             " receives the run's output artifacts")
-             f" {KeywordGroups.RESULT_KEYWORDS_TXT},"
-             f" {KeywordGroups.RESULT_GROUPS_CSV},"
-             f" {KeywordsToMerge.KEYWORDS_TO_MERGE_JSON},"
-             f" and {RunMeta.META_JSON}")
     parser.add_argument(
         "--model", default=model,
         help=f"the sentence embedding model (default \"{model}\")")
@@ -698,16 +687,9 @@ def parse_args(argv: list[str] | None) -> argparse.Namespace:
 def main(argv: list[str] | None = None) -> int:
     """Pool the two tagging runs' keywords and cluster them.
 
-    Writes the five fixed-named artifacts under the output
+    Creates the output directory (with parents) if it does not
-    directory, creating it (with parents) if it does not exist:
+    exist.  Each output artifact is written as soon as its
-    the pooled keyword text file; then the group membership CSV
+    content is computed, so when the input is rejected, or an
-    file, holding the clustering result alone; the group name
-    keyword text file, holding the same group names as a readable
-    list; the coding keyword set JSON file, holding the group
-    names plus every extra keyword given via ``--extra-keyword``;
-    and the run metadata JSON file, recording the command-line
-    choices and the environment.  Each file is written as soon as
-    its content is computed, so when the input is rejected, or an
     extra keyword duplicates a group name or another extra
     keyword, the output directory holds whatever the steps before
     the failing one produced, and the error message names what
@@ -719,8 +701,8 @@ def main(argv: list[str] | None = None) -> int:
     """
     started: float = time.monotonic()
     args: argparse.Namespace = parse_args(argv)
-    args.output_dir.mkdir(parents=True, exist_ok=True)
     try:
+        args.output_dir.mkdir(parents=True, exist_ok=True)
         source: PooledKeywords = KeywordPooler(
             args.run_dir_1, args.run_dir_2, args.output_dir).run()
         clusters: KeywordGroups = KeywordClusterer(
@@ -735,7 +717,7 @@ def main(argv: list[str] | None = None) -> int:
             f"Done.  Clustered {len(source.keywords)} keywords into"
             f" {len(clusters.names)}.  {elapsed} elapsed.",
             file=sys.stderr)
-    except ClusterError as error:
+    except (ClusterError, OSError) as error:
         print(f"error: {error}", file=sys.stderr)
         return 1
     return 0

tools/src/pop_fem_audit_tools/commands/export_llm_input.py

@@ -11,26 +11,17 @@ project's lyrics-only firewall: the output carries only the
 lyrics text of each song, identified by an opaque song key; no
 title, artist, or chart data crosses into the LLM input.
 
-With ``--extras``, each record's ``content`` becomes a JSON
+With ``--extras`` and ``--extras-per-id``, a record's content may
-object serialized as a string, its ``lyrics`` key holding the
+carry extra parameters alongside the lyrics, so a step that needs
-song's lyrics followed by the keys of the given extras file in
+them can get them without this module knowing what they mean; see
-their file order, so a step that needs parameters alongside the
+the exporter's content-building step for how the two merge.
-lyrics can carry them without this module knowing what they mean.
-
-With ``--extras-per-id``, the same merge happens per song: the
-given file maps a song ID to the extra keys of that one song, and
-the export is restricted to the song IDs the file names, so a step
-that revisits only some of the songs, each with its own parameters,
-gets exactly those records.  The two options may be given together,
-in which case a record's keys are ``lyrics``, the shared extras'
-keys, then that song's own keys, each group in its file order.
 """
 import argparse
 import json
 import sys
 import time
 from pathlib import Path
-from typing import Any
+from typing import Any, ClassVar
 
 import sqlalchemy as sa
 from sqlalchemy.orm import Session
@@ -40,6 +31,270 @@ from ..models import Song
 from ..utils import format_duration
 
 
+class LlmInputExporter:
+    """The exporter of the LLM input JSONL file."""
+
+    __LYRICS_KEY: ClassVar[str] = "lyrics"
+    """The key holding the lyrics in a record's merged content,
+    and the key forbidden in an extras file."""
+
+    def __init__(
+            self, output_jsonl: Path, extras: Path | None = None,
+            extras_per_id: Path | None = None) -> None:
+        """Set up the exporter of the LLM input JSONL file.
+
+        :param output_jsonl: The JSONL output file.
+        :param extras: The extras JSON file, or None for none.
+        :param extras_per_id: The per-ID extras JSON file, or
+            None for none.
+        """
+        self.__output_jsonl: Path = output_jsonl
+        """The JSONL output file."""
+        self.__extras_path: Path | None = extras
+        """The extras JSON file, or None for none."""
+        self.__extras_per_id_path: Path | None = extras_per_id
+        """The per-ID extras JSON file, or None for none."""
+
+    def run(self) -> int:
+        """Export the songs' lyrics to the output JSONL file.
+
+        :return: The number of songs exported.
+        :raises OSError: When a file cannot be read or written.
+        :raises sqlalchemy.exc.SQLAlchemyError: When the working
+            store cannot be read.
+        :raises ValueError: When an extras file is malformed, an
+            exported song has no lyrics, or the per-ID extras
+            name a song the working store does not have.
+        """
+        session: Session = ds.get_db()
+        try:
+            extras: dict[str, Any] | None = None
+            if self.__extras_path is not None:
+                extras = self.__load_extras(self.__extras_path)
+            extras_per_id: dict[str, dict[str, Any]] | None = None
+            if self.__extras_per_id_path is not None:
+                extras_per_id = self.__load_extras_per_id(
+                    self.__extras_per_id_path)
+            lines: list[str] = self.__build_lines(
+                session, extras, extras_per_id)
+        finally:
+            session.close()
+        self.__write_output(lines)
+        return len(lines)
+
+    @staticmethod
+    def __no_duplicate_keys(
+            pairs: list[tuple[str, Any]]) -> dict[str, Any]:
+        """Build a dict from JSON object pairs, rejecting
+        duplicates.
+
+        :param pairs: The key-value pairs of a JSON object, in
+            file order.
+        :return: The pairs as a dict, in file order.
+        :raises ValueError: When a key appears more than once.
+        """
+        result: dict[str, Any] = {}
+        key: str
+        value: Any
+        for key, value in pairs:
+            if key in result:
+                raise ValueError(
+                    f"duplicate key \"{key}\" in extras")
+            result[key] = value
+        return result
+
+    @classmethod
+    def __load_json_object(
+            cls, path: Path, label: str) -> dict[str, Any]:
+        """Load a single JSON object from a file, in file order.
+
+        :param path: The JSON file.
+        :param label: The kind of file, for the error messages.
+        :return: The object, in file order.
+        :raises OSError: When the file cannot be read.
+        :raises ValueError: When the file is not valid JSON, is
+            not a JSON object, or has duplicate keys.
+        """
+        with open(path, encoding="utf-8") as file:
+            text: str = file.read()
+        try:
+            data: Any = json.loads(
+                text, object_pairs_hook=cls.__no_duplicate_keys)
+        except json.JSONDecodeError as error:
+            raise ValueError(
+                f"invalid JSON in {label} file {path}: {error}") \
+                from error
+        if not isinstance(data, dict):
+            raise ValueError(
+                f"{label} file {path} must contain a JSON object")
+        return data
+
+    @classmethod
+    def __load_extras(cls, path: Path) -> dict[str, Any]:
+        """Load the extras object from a JSON file.
+
+        :param path: The extras JSON file.
+        :return: The extras, in file order.
+        :raises OSError: When the file cannot be read.
+        :raises ValueError: When the file is not valid JSON, is
+            not a JSON object, has duplicate keys, or has a
+            "lyrics" key.
+        """
+        data: dict[str, Any] = cls.__load_json_object(
+            path, "extras")
+        if cls.__LYRICS_KEY in data:
+            raise ValueError(
+                f"extras file {path} must not have a"
+                f" \"{cls.__LYRICS_KEY}\" key")
+        return data
+
+    @classmethod
+    def __load_extras_per_id(
+            cls, path: Path) -> dict[str, dict[str, Any]]:
+        """Load the per-ID extras object from a JSON file.
+
+        :param path: The per-ID extras JSON file, mapping a song
+            ID, as ``song-<N>``, to the extras of that one song.
+        :return: The extras of each song ID, in file order, every
+            song's own extras in their file order too.
+        :raises OSError: When the file cannot be read.
+        :raises ValueError: When the file is not valid JSON, is
+            not a JSON object, has duplicate keys, has a song
+            whose value is not a JSON object, or has a song with
+            a "lyrics" key.
+        """
+        data: dict[str, Any] = cls.__load_json_object(
+            path, "per-ID extras")
+        song_id: str
+        extras: Any
+        for song_id, extras in data.items():
+            if not isinstance(extras, dict):
+                raise ValueError(
+                    f"per-ID extras file {path}: id {song_id}"
+                    " must have a JSON object")
+            if cls.__LYRICS_KEY in extras:
+                raise ValueError(
+                    f"per-ID extras file {path}: id {song_id}"
+                    f" must not have a \"{cls.__LYRICS_KEY}\""
+                    " key")
+        return data
+
+    def __build_lines(
+            self, session: Session,
+            extras: dict[str, Any] | None = None,
+            extras_per_id: dict[str, dict[str, Any]] | None
+            = None) -> list[str]:
+        """Build the JSONL lines of the exported songs' lyrics.
+
+        Every song is exported, unless per-ID extras are given,
+        in which case only the songs they name are; see
+        :meth:`__build_content` for how the extras merge into a
+        record's content.
+
+        :param session: The database session.
+        :param extras: The extra parameters merged into every
+            record's content alongside the lyrics, in the order
+            they are to appear, or None for none.
+        :param extras_per_id: The extra parameters merged into
+            the content of one record alone, keyed by that
+            record's song ID and in the order they are to appear,
+            restricting the export to the song IDs they name, or
+            None for no such extras and no such restriction.
+        :return: The JSON lines, one per exported song, ordered by
+            song ID.
+        :raises ValueError: When an exported song has no lyrics,
+            or the per-ID extras name a song the working store
+            does not have.
+        """
+        lines: list[str] = []
+        exported: set[str] = set()
+        song: Song
+        for song in session.scalars(
+                sa.select(Song).order_by(Song.id)):
+            song_id: str = f"song-{song.id}"
+            if extras_per_id is not None \
+                    and song_id not in extras_per_id:
+                continue
+            if song.lyrics is None:
+                raise ValueError(
+                    f"song {song.id} \"{song.title}\": no lyrics")
+            song_extras: dict[str, Any] | None = None \
+                if extras_per_id is None \
+                else extras_per_id[song_id]
+            content: str = self.__build_content(
+                song.lyrics, extras, song_extras)
+            record: dict[str, str] = {
+                "id": song_id, "content": content}
+            lines.append(json.dumps(record, ensure_ascii=False))
+            exported.add(song_id)
+        if extras_per_id is not None:
+            missing: list[str] = sorted(
+                set(extras_per_id) - exported)
+            if len(missing) > 0:
+                raise ValueError(
+                    "the per-ID extras name songs the working"
+                    f" store does not have: {', '.join(missing)}")
+        return lines
+
+    @classmethod
+    def __build_content(
+            cls, lyrics: str, extras: dict[str, Any] | None,
+            song_extras: dict[str, Any] | None) -> str:
+        """Build the content of one exported record.
+
+        Without extras of either kind, a record's content is the
+        bare lyrics string.  With ``--extras``, the content
+        becomes a JSON object serialized as a string, its
+        "lyrics" key holding the song's lyrics followed by the
+        keys of the given extras file, in their file order.  With
+        ``--extras-per-id``, the same merge happens per song: the
+        song's own extra keys follow the lyrics instead.  When
+        both are given, a record's keys are "lyrics", the shared
+        extras' keys, then that song's own keys, each group in
+        its file order.
+
+        :param lyrics: The lyrics of the song.
+        :param extras: The extra parameters shared by every
+            record, in the order they are to appear, or None for
+            none.
+        :param song_extras: The extra parameters of this record
+            alone, in the order they are to appear, or None for
+            none.
+        :return: The bare lyrics when there are no extras of
+            either kind, or otherwise a JSON object serialized as
+            a string, whose first key is "lyrics" holding the
+            lyrics, followed by the shared extras' keys and then
+            this record's own keys, each group in its given
+            order.
+        """
+        if extras is None and song_extras is None:
+            return lyrics
+        payload: dict[str, Any] = {cls.__LYRICS_KEY: lyrics}
+        if extras is not None:
+            payload.update(extras)
+        if song_extras is not None:
+            payload.update(song_extras)
+        return json.dumps(payload, ensure_ascii=False)
+
+    def __write_output(self, lines: list[str]) -> None:
+        """Write the exported lines to the output JSONL file.
+
+        Creates the parent directory when it does not exist.
+
+        :param lines: The JSONL lines, in the output order.
+        :return: None.
+        :raises OSError: When the file cannot be written.
+        """
+        self.__output_jsonl.parent.mkdir(
+            parents=True, exist_ok=True)
+        with open(
+                self.__output_jsonl, "w",
+                encoding="utf-8") as file:
+            line: str
+            for line in lines:
+                file.write(line + "\n")
+
+
 def parse_args(argv: list[str] | None) -> argparse.Namespace:
     """Parse the command-line arguments.
 
@@ -56,227 +311,33 @@ def parse_args(argv: list[str] | None) -> argparse.Namespace:
     parser.add_argument(
         "--extras", type=Path, default=None,
         help="a JSON file holding a single JSON object of extra"
-             " parameters; when given, each record's \"content\""
+             " parameters merged into every record's content")
-             " becomes a JSON object string with a \"lyrics\" key"
-             " followed by the extras' keys, instead of the bare"
-             " lyrics string")
     parser.add_argument(
         "--extras-per-id", type=Path, default=None,
         help="a JSON file holding a single JSON object that maps a"
              " song ID, as \"song-<N>\", to a JSON object of extra"
-             " parameters for that one song; the song's object is"
+             " parameters for that one song, restricting the"
-             " merged into its \"content\" the same way as with"
+             " export to the song IDs the file names")
-             " --extras, and the export is restricted to the song"
-             " IDs the file names")
     return parser.parse_args(argv)
 
 
-def __no_duplicate_keys(
-        pairs: list[tuple[str, Any]]) -> dict[str, Any]:
-    """Build a dict from JSON object pairs, rejecting duplicates.
-
-    :param pairs: The key-value pairs of a JSON object, in file
-        order.
-    :return: The pairs as a dict, in file order.
-    :raises ValueError: When a key appears more than once.
-    """
-    result: dict[str, Any] = {}
-    key: str
-    value: Any
-    for key, value in pairs:
-        if key in result:
-            raise ValueError(f"duplicate key \"{key}\" in extras")
-        result[key] = value
-    return result
-
-
-def __load_json_object(path: Path, label: str) -> dict[str, Any]:
-    """Load a single JSON object from a file, in file order.
-
-    :param path: The JSON file.
-    :param label: The kind of file, for the error messages.
-    :return: The object, in file order.
-    :raises OSError: When the file cannot be read.
-    :raises ValueError: When the file is not valid JSON, is not
-        a JSON object, or has duplicate keys.
-    """
-    with open(path, encoding="utf-8") as file:
-        text: str = file.read()
-    try:
-        data: Any = json.loads(
-            text, object_pairs_hook=__no_duplicate_keys)
-    except json.JSONDecodeError as error:
-        raise ValueError(
-            f"invalid JSON in {label} file {path}: {error}") \
-            from error
-    if not isinstance(data, dict):
-        raise ValueError(
-            f"{label} file {path} must contain a JSON object")
-    return data
-
-
-def load_extras(path: Path) -> dict[str, Any]:
-    """Load the extras object from a JSON file.
-
-    :param path: The extras JSON file.
-    :return: The extras, in file order.
-    :raises OSError: When the file cannot be read.
-    :raises ValueError: When the file is not valid JSON, is not
-        a JSON object, has duplicate keys, or has a "lyrics" key.
-    """
-    data: dict[str, Any] = __load_json_object(path, "extras")
-    if "lyrics" in data:
-        raise ValueError(
-            f"extras file {path} must not have a \"lyrics\" key")
-    return data
-
-
-def load_extras_per_id(path: Path) -> dict[str, dict[str, Any]]:
-    """Load the per-ID extras object from a JSON file.
-
-    :param path: The per-ID extras JSON file, mapping a song ID,
-        as ``song-<N>``, to the extras of that one song.
-    :return: The extras of each song ID, in file order, every
-        song's own extras in their file order too.
-    :raises OSError: When the file cannot be read.
-    :raises ValueError: When the file is not valid JSON, is not
-        a JSON object, has duplicate keys, has a song whose value
-        is not a JSON object, or has a song with a "lyrics" key.
-    """
-    data: dict[str, Any] = __load_json_object(path, "per-ID extras")
-    song_id: str
-    extras: Any
-    for song_id, extras in data.items():
-        if not isinstance(extras, dict):
-            raise ValueError(
-                f"per-ID extras file {path}: id {song_id} must"
-                " have a JSON object")
-        if "lyrics" in extras:
-            raise ValueError(
-                f"per-ID extras file {path}: id {song_id} must not"
-                " have a \"lyrics\" key")
-    return data
-
-
-def __build_content(
-        lyrics: str,
-        extras: dict[str, Any] | None,
-        song_extras: dict[str, Any] | None) -> str:
-    """Build the content of one exported record.
-
-    :param lyrics: The lyrics of the song.
-    :param extras: The extra parameters shared by every record,
-        in the order they are to appear, or None for none.
-    :param song_extras: The extra parameters of this record
-        alone, in the order they are to appear, or None for none.
-    :return: The bare lyrics when there are no extras of either
-        kind, or otherwise a JSON object serialized as a string,
-        whose first key is ``"lyrics"`` holding the lyrics,
-        followed by the shared extras' keys and then this
-        record's own keys, each group in its given order.
-    """
-    if extras is None and song_extras is None:
-        return lyrics
-    payload: dict[str, Any] = {"lyrics": lyrics}
-    if extras is not None:
-        payload.update(extras)
-    if song_extras is not None:
-        payload.update(song_extras)
-    return json.dumps(payload, ensure_ascii=False)
-
-
-def build_lines(
-        session: Session,
-        extras: dict[str, Any] | None = None,
-        extras_per_id: dict[str, dict[str, Any]] | None = None) \
-        -> list[str]:
-    """Build the JSONL lines of the exported songs' lyrics.
-
-    Without extras of either kind, each record's ``content`` is
-    the bare lyrics string.  With extras, ``content`` is a JSON
-    object serialized as a string, whose first key is ``"lyrics"``
-    holding the lyrics string, followed by the shared extras' keys
-    and then the song's own per-ID extras' keys, each group in its
-    given order.
-
-    Every song is exported, unless per-ID extras are given, in
-    which case only the songs they name are.
-
-    :param session: The database session.
-    :param extras: The extra parameters merged into every
-        record's content alongside the lyrics, in the order they
-        are to appear, or None for none.
-    :param extras_per_id: The extra parameters merged into the
-        content of one record alone, keyed by that record's song
-        ID and in the order they are to appear, restricting the
-        export to the song IDs they name, or None for no such
-        extras and no such restriction.
-    :return: The JSON lines, one per exported song, ordered by
-        song ID.
-    :raises ValueError: When an exported song has no lyrics, or
-        the per-ID extras name a song the working store does not
-        have.
-    """
-    lines: list[str] = []
-    exported: set[str] = set()
-    song: Song
-    for song in session.scalars(sa.select(Song).order_by(Song.id)):
-        song_id: str = f"song-{song.id}"
-        if extras_per_id is not None and song_id not in extras_per_id:
-            continue
-        if song.lyrics is None:
-            raise ValueError(
-                f"song {song.id} \"{song.title}\": no lyrics")
-        song_extras: dict[str, Any] | None = None \
-            if extras_per_id is None else extras_per_id[song_id]
-        content: str = __build_content(
-            song.lyrics, extras, song_extras)
-        record: dict[str, str] = {
-            "id": song_id, "content": content}
-        lines.append(json.dumps(record, ensure_ascii=False))
-        exported.add(song_id)
-    if extras_per_id is not None:
-        missing: list[str] = sorted(set(extras_per_id) - exported)
-        if len(missing) > 0:
-            raise ValueError(
-                "the per-ID extras name songs the working store"
-                f" does not have: {', '.join(missing)}")
-    return lines
-
-
 def main(argv: list[str] | None = None) -> int:
     """Export the LLM input JSONL file from the working store.
 
-    Every song is exported, unless ``--extras-per-id`` is given,
-    in which case only the songs its file names are.
-
     :param argv: The command-line arguments, or None for
         ``sys.argv``.
     :return: The exit status: 0 on success, non-zero on failure.
     """
     started: float = time.monotonic()
     args: argparse.Namespace = parse_args(argv)
-    session: Session = ds.get_db()
-    lines: list[str]
     try:
-        extras: dict[str, Any] | None = None
+        count: int = LlmInputExporter(
-        if args.extras is not None:
+            args.output_jsonl, args.extras,
-            extras = load_extras(args.extras)
+            args.extras_per_id).run()
-        extras_per_id: dict[str, dict[str, Any]] | None = None
-        if args.extras_per_id is not None:
-            extras_per_id = load_extras_per_id(args.extras_per_id)
-        lines = build_lines(session, extras, extras_per_id)
     except (OSError, sa.exc.SQLAlchemyError, ValueError) as error:
         print(f"error: {error}", file=sys.stderr)
         return 1
-    finally:
-        session.close()
-    args.output_jsonl.parent.mkdir(parents=True, exist_ok=True)
-    with open(args.output_jsonl, "w", encoding="utf-8") as file:
-        line: str
-        for line in lines:
-            file.write(line + "\n")
     elapsed: str = format_duration(time.monotonic() - started)
-    print(f"Done.  {len(lines)} songs exported."
+    print(f"Done.  {count} songs exported."
           f"  {elapsed} elapsed.", file=sys.stderr)
     return 0

tools/src/pop_fem_audit_tools/commands/fetch_artists.py

@@ -12,12 +12,10 @@ layer.  The working store is only read, never written; the
 ``build-db`` subcommand assembles the captured files into the
 store on the next rebuild.
 
-Every fetched row is meant for later human verification: the
+An unresolved artist or an error on one artist is noted on its
-description of the resolved item is recorded in the note column
+row and does not fail the run.  A row whose name is no longer an
-so that a bad match can be spotted.  An unresolved artist or an
+artist of the store is dropped from the snapshot and reported on
-error on one artist is noted on its row and does not fail the
+the standard error.
-run.  A row whose name is no longer an artist of the store is
-dropped from the snapshot and reported on the standard error.
 """
 import argparse
 import csv
@@ -33,7 +31,7 @@ import urllib.request
 from collections.abc import Container, Sequence
 from dataclasses import asdict, dataclass, field, fields
 from pathlib import Path
-from typing import Any, Literal, TextIO
+from typing import Any, ClassVar, Literal, TextIO
 
 import sqlalchemy as sa
 from sqlalchemy.orm import Session
@@ -43,71 +41,6 @@ from ..database import ds
 from ..models import Artist, Song, SongArtist
 from ..utils import format_duration
 
-API_URL: str = "https://www.wikidata.org/w/api.php"
-"""The URL of the Wikidata API endpoint."""
-SPARQL_URL: str = "https://query.wikidata.org/sparql"
-"""The URL of the Wikidata Query Service SPARQL endpoint."""
-USER_AGENT: str = (
-    f"pop-fem-audit-tools/{VERSION}"
-    " (https://github.com/imacat/pop-fem-audit;"
-    " mailto:imacat@mail.imacat.idv.tw)")
-"""The User-Agent header sent on every HTTP request."""
-TIMEOUT: float = 30.0
-"""The timeout of an API HTTP request, in seconds."""
-SPARQL_TIMEOUT: float = 90.0
-"""The timeout of a SPARQL HTTP request, in seconds.
-
-Higher than the API timeout: the WDQS server aborts a slow
-query at 60 seconds, and a lower client timeout would race
-that server-side abort and misclassify a slow-but-answerable
-query as a client-side timeout instead of letting the server's
-own HTTP error response arrive and enter the retry path."""
-SLEEP_SECONDS: float = 1.0
-"""The delay between consecutive HTTP requests, in seconds."""
-MAX_ATTEMPTS: int = 5
-"""The maximum number of attempts on a transient error."""
-RETRY_SECONDS: float = 15.0
-"""The back-off unit on a transient error, in seconds;
-multiplied by the attempt number already made."""
-RETRY_STATUSES: frozenset[int] = frozenset({429, 500, 502, 503})
-"""The HTTP statuses that are retried with a back-off."""
-MAX_STAGE1_TITLES: int = 3
-"""The maximum number of charted titles used for the stage-1 song
-corroboration."""
-HUMAN_QID: str = "Q5"
-"""The Wikidata item ID of "human"."""
-ENSEMBLE_QID: str = "Q2088357"
-"""The Wikidata item ID of "musical ensemble"."""
-ORIGINAL_CAST_QID: str = "Q106497009"
-"""The Wikidata item ID of "original cast"."""
-GROUP_KEYWORDS: Sequence[str] = ("band", "group", "duo", "trio")
-"""The label keywords that suggest a musical ensemble, covering
-labels like "boy band" and "girl group"."""
-NOTE_NOT_FOUND: str = "not found"
-"""The note sentinel of an artist without a resolved Wikidata
-item, written to the snapshot and read back for the
-classification."""
-CORPUS_START_YEAR: int = 2016
-"""The first year of the corpus window: a member who left a group
-before it never performed a corpus song."""
-MIXED_GENDER: str = "mixed"
-"""The gender recorded for a group whose members do not share one
-gender."""
-TIME_YEAR_PATTERN: re.Pattern[str] = re.compile(r"^[+-]?\d+")
-"""The leading year of a Wikidata time value."""
-PINNED_QIDS: dict[str, str] = {}
-"""The last-resort pinned item IDs, keyed by the artist name.
-
-An entry is for an artist the algorithm documented on
-``ArtistFetcher`` is structurally unable to resolve, with its
-justification recorded here.  Currently empty: the only pin ever
-needed, "Pinkfong" (typed as a brand, which the type gate
-excludes by design), became moot when the store's artist entity
-behind that credit was identified as Hope Segoine.
-
-A pinned name skips the candidate retrieval and corroboration
-steps; its item ID is used directly."""
-
 
 class ArtistType(enum.StrEnum):
     """The decided artist type of a snapshot row."""
@@ -147,15 +80,14 @@ class ArtistSnapshot:
         return asdict(self)
 
 
-SNAPSHOT_FIELDS: Sequence[str] = tuple(
-    x.name for x in fields(ArtistSnapshot))
-"""The header columns of the Wikidata artist snapshot CSV file."""
-
-
 @dataclass
 class GroupMember:
     """One has-part member of a Wikidata group item."""
 
+    __CORPUS_START_YEAR: ClassVar[int] = 2016
+    """The first year of the corpus window: a member who left a
+    group before it never performed a corpus song."""
+
     qid: str
     """The item ID of the member."""
     start_years: list[int] = field(default_factory=list)
@@ -181,7 +113,7 @@ class GroupMember:
         if len(self.end_years) == 0:
             return True
         last_end: int = max(self.end_years)
-        if last_end >= CORPUS_START_YEAR:
+        if last_end >= self.__CORPUS_START_YEAR:
             return True
         if len(self.start_years) == 0:
             return False
@@ -229,20 +161,8 @@ class RetryExhausted(Exception):
     """
 
 
-def parse_args(argv: list[str] | None) -> argparse.Namespace:
+NOTE_NOT_FOUND: str = "not found"
-    """Parse the command-line arguments.
+"""The note marking an artist that could not be resolved."""
-
-    :param argv: The command-line arguments, or None for
-        ``sys.argv``.
-    :return: The parsed arguments.
-    """
-    parser: argparse.ArgumentParser = argparse.ArgumentParser(
-        description="Fetch the artist metadata from Wikidata"
-                    " into the capture layer.")
-    parser.add_argument(
-        "wikidata_csv", type=Path,
-        help="the Wikidata artist snapshot CSV file")
-    return parser.parse_args(argv)
 
 
 class ArtistFetcher:
@@ -291,6 +211,71 @@ class ArtistFetcher:
     in the note.
     """
 
+    __API_URL: ClassVar[str] = "https://www.wikidata.org/w/api.php"
+    """The URL of the Wikidata API endpoint."""
+    __SPARQL_URL: ClassVar[str] \
+        = "https://query.wikidata.org/sparql"
+    """The URL of the Wikidata Query Service SPARQL endpoint."""
+    __USER_AGENT: ClassVar[str] = (
+        f"pop-fem-audit-tools/{VERSION}"
+        " (https://github.com/imacat/pop-fem-audit;"
+        " mailto:imacat@mail.imacat.idv.tw)")
+    """The User-Agent header sent on every HTTP request."""
+    __TIMEOUT: ClassVar[float] = 30.0
+    """The timeout of an API HTTP request, in seconds."""
+    __SPARQL_TIMEOUT: ClassVar[float] = 90.0
+    """The timeout of a SPARQL HTTP request, in seconds.
+
+    Higher than the API timeout: the WDQS server aborts a slow
+    query at 60 seconds, and a lower client timeout would race
+    that server-side abort and misclassify a slow-but-answerable
+    query as a client-side timeout instead of letting the
+    server's own HTTP error response arrive and enter the retry
+    path."""
+    __SLEEP_SECONDS: ClassVar[float] = 1.0
+    """The delay between consecutive HTTP requests, in
+    seconds."""
+    __MAX_ATTEMPTS: ClassVar[int] = 5
+    """The maximum number of attempts on a transient error."""
+    __RETRY_SECONDS: ClassVar[float] = 15.0
+    """The back-off unit on a transient error, in seconds;
+    multiplied by the attempt number already made."""
+    __RETRY_STATUSES: ClassVar[frozenset[int]] \
+        = frozenset({429, 500, 502, 503})
+    """The HTTP statuses that are retried with a back-off."""
+    __MAX_STAGE1_TITLES: ClassVar[int] = 3
+    """The maximum number of charted titles used for the
+    stage-1 song corroboration."""
+    __HUMAN_QID: ClassVar[str] = "Q5"
+    """The Wikidata item ID of "human"."""
+    __ENSEMBLE_QID: ClassVar[str] = "Q2088357"
+    """The Wikidata item ID of "musical ensemble"."""
+    __ORIGINAL_CAST_QID: ClassVar[str] = "Q106497009"
+    """The Wikidata item ID of "original cast"."""
+    __GROUP_KEYWORDS: ClassVar[Sequence[str]] \
+        = ("band", "group", "duo", "trio")
+    """The label keywords that suggest a musical ensemble,
+    covering labels like "boy band" and "girl group"."""
+    __MIXED_GENDER: ClassVar[str] = "mixed"
+    """The gender recorded for a group whose members do not
+    share one gender."""
+    __TIME_YEAR_PATTERN: ClassVar[re.Pattern[str]] \
+        = re.compile(r"^[+-]?\d+")
+    """The leading year of a Wikidata time value."""
+    __PINNED_QIDS: ClassVar[dict[str, str]] = {}
+    """The last-resort pinned item IDs, keyed by the artist name.
+
+    An entry is for an artist the algorithm documented on
+    ``ArtistFetcher`` is structurally unable to resolve, with its
+    justification recorded here.  Currently empty: the only pin
+    ever needed, "Pinkfong" (typed as a brand, which the type
+    gate excludes by design), became moot when the store's
+    artist entity behind that credit was identified as Hope
+    Segoine.
+
+    A pinned name skips the candidate retrieval and corroboration
+    steps; its item ID is used directly."""
+
     def __init__(self) -> None:
         """Construct the fetcher."""
         self.__sent: int = 0
@@ -342,8 +327,8 @@ class ArtistFetcher:
             transient error are exhausted.
         :raises ValueError: On a JSON decoding error.
         """
-        if name in PINNED_QIDS:
+        if name in self.__PINNED_QIDS:
-            return PINNED_QIDS[name]
+            return self.__PINNED_QIDS[name]
         candidates: list[str] = self.__candidates(name)
         if len(candidates) == 0:
             return None
@@ -373,11 +358,11 @@ class ArtistFetcher:
           {{ ?item rdfs:label ?name }}
           UNION {{ ?item skos:altLabel ?name }}
           {{
-            ?item wdt:P31 wd:{HUMAN_QID}
+            ?item wdt:P31 wd:{self.__HUMAN_QID}
           }} UNION {{
-            ?item wdt:P31/wdt:P279* wd:{ENSEMBLE_QID}
+            ?item wdt:P31/wdt:P279* wd:{self.__ENSEMBLE_QID}
           }} UNION {{
-            ?item wdt:P31 wd:{ORIGINAL_CAST_QID}
+            ?item wdt:P31 wd:{self.__ORIGINAL_CAST_QID}
           }}
         }}
         """
@@ -403,7 +388,7 @@ class ArtistFetcher:
             transient error are exhausted.
         :raises ValueError: On a JSON decoding error.
         """
-        subset: Sequence[str] = titles[:MAX_STAGE1_TITLES]
+        subset: Sequence[str] = titles[:self.__MAX_STAGE1_TITLES]
         if len(subset) == 0:
             return None
         query: str = f"""
@@ -532,7 +517,7 @@ class ArtistFetcher:
         qid: str
         for qid in qids:
             member: MemberClaims = claims.get(qid, MemberClaims())
-            if HUMAN_QID not in member.instance_of_ids:
+            if self.__HUMAN_QID not in member.instance_of_ids:
                 continue
             if len(member.gender_ids) == 0:
                 return
@@ -543,7 +528,7 @@ class ArtistFetcher:
             [x[1] for x in genders], any_language=True)
         unique: set[str] = {x[1] for x in genders}
         snapshot.gender = labels[genders[0][1]] \
-            if len(unique) == 1 else MIXED_GENDER
+            if len(unique) == 1 else self.__MIXED_GENDER
         basis: str = "gender derived from members: " + "; ".join(
             f"{x} {labels[y]}" for x, y in genders)
         snapshot.note = f"{snapshot.note}; {basis}" \
@@ -677,7 +662,8 @@ class ArtistFetcher:
                     or not isinstance(value.get("time"), str):
                 continue
             match: re.Match[str] | None \
-                = TIME_YEAR_PATTERN.match(value["time"])
+                = ArtistFetcher.__TIME_YEAR_PATTERN.match(
+                    value["time"])
             if match is not None:
                 years.append(int(match.group()))
         return years
@@ -806,12 +792,13 @@ class ArtistFetcher:
             ``ArtistType.GROUP`` for a musical ensemble, or the
             empty string for the human to decide.
         """
-        if HUMAN_QID in type_ids:
+        if ArtistFetcher.__HUMAN_QID in type_ids:
             return ArtistType.SOLO
         qid: str
         for qid in type_ids:
             label: str = labels.get(qid, "").lower()
-            if any(x in label for x in GROUP_KEYWORDS):
+            if any(x in label
+                   for x in ArtistFetcher.__GROUP_KEYWORDS):
                 return ArtistType.GROUP
         return ""
 
@@ -827,14 +814,15 @@ class ArtistFetcher:
             transient error are exhausted.
         :raises ValueError: On a JSON decoding error.
         """
-        url: str = (f"{SPARQL_URL}?"
+        url: str = (
-                    f"{urllib.parse.urlencode({'query': query})}")
+            f"{self.__SPARQL_URL}?"
+            f"{urllib.parse.urlencode({'query': query})}")
         request: urllib.request.Request = urllib.request.Request(
             url, headers={
-                "User-Agent": USER_AGENT,
+                "User-Agent": self.__USER_AGENT,
                 "Accept": "application/sparql-results+json"})
         body: bytes = self.__send(
-            request, timeout=SPARQL_TIMEOUT)
+            request, timeout=self.__SPARQL_TIMEOUT)
         data: Any = json.loads(body)
         bindings: Any = None
         if isinstance(data, dict) \
@@ -868,13 +856,14 @@ class ArtistFetcher:
             transient error are exhausted.
         :raises ValueError: On a JSON decoding error.
         """
-        url: str = f"{API_URL}?{urllib.parse.urlencode(params)}"
+        url: str \
+            = f"{self.__API_URL}?{urllib.parse.urlencode(params)}"
         request: urllib.request.Request = urllib.request.Request(
-            url, headers={"User-Agent": USER_AGENT})
+            url, headers={"User-Agent": self.__USER_AGENT})
         return json.loads(self.__send(request))
 
     def __send(self, request: urllib.request.Request,
-               timeout: float = TIMEOUT) -> bytes:
+               timeout: float = __TIMEOUT) -> bytes:
         """Send an HTTP request, retrying on a transient error.
 
         Consecutive requests are separated by a fixed delay.  A
@@ -892,7 +881,7 @@ class ArtistFetcher:
             transient error are exhausted.
         """
         if self.__sent > 0:
-            time.sleep(SLEEP_SECONDS)
+            time.sleep(self.__SLEEP_SECONDS)
         self.__sent += 1
         attempt: int = 1
         reason: str | None
@@ -905,10 +894,10 @@ class ArtistFetcher:
                 reason = self.__retry_reason(error)
                 if reason is None:
                     raise
-                if attempt >= MAX_ATTEMPTS:
+                if attempt >= self.__MAX_ATTEMPTS:
                     raise RetryExhausted(
                         f"retries exhausted ({reason})") from error
-            time.sleep(RETRY_SECONDS * attempt)
+            time.sleep(self.__RETRY_SECONDS * attempt)
             attempt += 1
 
     @staticmethod
@@ -923,7 +912,7 @@ class ArtistFetcher:
             the error is not transient and must not be retried.
         """
         if isinstance(error, urllib.error.HTTPError):
-            if error.code not in RETRY_STATUSES:
+            if error.code not in ArtistFetcher.__RETRY_STATUSES:
                 return None
             return str(error)
         if isinstance(error, TimeoutError):
@@ -968,92 +957,216 @@ class ArtistFetcher:
         return uri.rsplit("/", 1)[-1]
 
 
-def read_snapshot_rows(file: TextIO) -> list[dict[str, str]]:
+@dataclass(frozen=True)
-    """Read the current rows of a snapshot CSV file handle.
+class FetchCounts:
+    """The outcome counts of one snapshot update run."""
 
-    :param file: The open, seekable snapshot CSV file.
+    fetched: int
-    :return: The rows, keyed by the column name.
+    """The number of artists newly resolved."""
-    :raises OSError: When the file cannot be read.
+    not_found: int
+    """The number of artists left unresolved."""
+    errors: int
+    """The number of artists that ended in an error."""
+
+
+class ArtistSnapshotUpdater:
+    """The updater of the Wikidata artist snapshot CSV file.
+
+    Fetches the metadata of every artist of the working store
+    that the snapshot does not resolve yet, appends a row for
+    each to the snapshot as it is fetched, and rewrites the
+    snapshot sorted by artist name with its stale rows dropped.
     """
-    file.seek(0)
-    reader: csv.DictReader[str] = csv.DictReader(file)
-    return list(reader)
 
+    __SNAPSHOT_FIELDS: ClassVar[Sequence[str]] = tuple(
+        x.name for x in fields(ArtistSnapshot))
+    """The header columns of the Wikidata artist snapshot CSV file."""
 
-def read_artist_titles(session: Session,
+    def __init__(self, wikidata_csv: Path) -> None:
-                       artist_id: int) -> list[str]:
+        """Set up the updater.
-    """Read the charted song titles credited to an artist.
 
-    :param session: The database session.
+        :param wikidata_csv: The Wikidata artist snapshot CSV
-    :param artist_id: The artist ID.
+            file.
-    :return: The song titles credited to the artist, ordered by
+        """
-        the song ID, with the duplicate titles removed.
+        self.__wikidata_csv: Path = wikidata_csv
-    """
+        """The Wikidata artist snapshot CSV file."""
-    titles: Sequence[str] = session.scalars(
-        sa.select(Song.title)
-        .join(SongArtist, SongArtist.song_id == Song.id)
-        .where(SongArtist.artist_id == artist_id)
-        .order_by(Song.id)).all()
-    return list(dict.fromkeys(titles))
 
+    def run(self) -> FetchCounts:
+        """Fetch every unresolved artist and update the snapshot.
 
-def ensure_snapshot_header(file: TextIO) -> None:
+        :return: The counts of the run.
-    """Write the snapshot CSV header row if the file is empty.
+        :raises OSError: When the snapshot file, or its parent
+            directory, cannot be read or written.
+        :raises sqlalchemy.exc.SQLAlchemyError: When the working
+            store cannot be read.
+        """
+        session: Session = ds.get_db()
+        try:
+            return self.__run(session)
+        finally:
+            session.close()
 
-    :param file: The open, seekable snapshot CSV file.
+    def __run(self, session: Session) -> FetchCounts:
-    :return: None.
+        """Run the fetch loop with an open database session.
-    :raises OSError: When the file cannot be written.
+
-    """
+        :param session: The database session.
-    file.seek(0, os.SEEK_END)
+        :return: The counts of the run.
-    if file.tell() == 0:
+        :raises OSError: When the snapshot file, or its parent
-        csv.writer(file).writerow(SNAPSHOT_FIELDS)
+            directory, cannot be read or written.
+        """
+        fetcher: ArtistFetcher = ArtistFetcher()
+        fetched: int = 0
+        not_found: int = 0
+        errors: int = 0
+        self.__wikidata_csv.parent.mkdir(
+            parents=True, exist_ok=True)
+        with open(self.__wikidata_csv, "a+", encoding="utf-8",
+                  newline="") as csv_file:
+            done: set[str] = {
+                x["name"] for x in
+                self.__read_snapshot_rows(csv_file)
+                if x["gender"] != ""}
+            self.__ensure_snapshot_header(csv_file)
+            names: set[str] = set()
+            artist: Artist
+            for artist in session.scalars(
+                    sa.select(Artist).order_by(Artist.id)):
+                names.add(artist.name)
+                if artist.name in done:
+                    continue
+                titles: list[str] = self.__read_artist_titles(
+                    session, artist.id)
+                snapshot: ArtistSnapshot = fetcher.fetch(
+                    artist.name, titles)
+                self.__append_row(csv_file, snapshot)
+                status: str = snapshot.qid
+                if snapshot.note == NOTE_NOT_FOUND:
+                    not_found += 1
+                    status = NOTE_NOT_FOUND
+                elif snapshot.note.startswith("error: "):
+                    errors += 1
+                    status = snapshot.note
+                else:
+                    fetched += 1
+                print(f"artist \"{artist.name}\": {status}",
+                      file=sys.stderr)
+            self.__write_snapshot(csv_file, names)
+        return FetchCounts(
+            fetched=fetched, not_found=not_found, errors=errors)
+
+    @staticmethod
+    def __read_snapshot_rows(file: TextIO) \
+            -> list[dict[str, str]]:
+        """Read the current rows of a snapshot CSV file handle.
+
+        :param file: The open, seekable snapshot CSV file.
+        :return: The rows, keyed by the column name.
+        :raises OSError: When the file cannot be read.
+        """
+        file.seek(0)
+        reader: csv.DictReader[str] = csv.DictReader(file)
+        return list(reader)
+
+    @staticmethod
+    def __read_artist_titles(session: Session,
+                             artist_id: int) -> list[str]:
+        """Read the charted song titles credited to an artist.
+
+        :param session: The database session.
+        :param artist_id: The artist ID.
+        :return: The song titles credited to the artist, ordered
+            by the song ID, with the duplicate titles removed.
+        """
+        titles: Sequence[str] = session.scalars(
+            sa.select(Song.title)
+            .join(SongArtist, SongArtist.song_id == Song.id)
+            .where(SongArtist.artist_id == artist_id)
+            .order_by(Song.id)).all()
+        return list(dict.fromkeys(titles))
+
+    @staticmethod
+    def __ensure_snapshot_header(file: TextIO) -> None:
+        """Write the snapshot CSV header row if the file is
+        empty.
+
+        :param file: The open, seekable snapshot CSV file.
+        :return: None.
+        :raises OSError: When the file cannot be written.
+        """
+        file.seek(0, os.SEEK_END)
+        if file.tell() == 0:
+            csv.writer(file).writerow(
+                ArtistSnapshotUpdater.__SNAPSHOT_FIELDS)
+            file.flush()
+
+    @staticmethod
+    def __append_row(file: TextIO,
+                     snapshot: ArtistSnapshot) -> None:
+        """Append a snapshot row to a snapshot CSV file handle.
+
+        :param file: The open snapshot CSV file, opened for
+            append.
+        :param snapshot: The snapshot of an artist.
+        :return: None.
+        :raises OSError: When the file cannot be written.
+        """
+        csv.DictWriter(
+            file, ArtistSnapshotUpdater.__SNAPSHOT_FIELDS).writerow(
+            snapshot.to_row())
         file.flush()
 
+    @staticmethod
+    def __write_snapshot(file: TextIO, names: Container[str]) \
+            -> None:
+        """Rewrite a snapshot CSV file handle sorted by artist
+        name.
 
-def append_row(file: TextIO, snapshot: ArtistSnapshot) -> None:
+        The rows are ordered by the case-folded artist name,
-    """Append a snapshot row to a snapshot CSV file handle.
+        matching the convention of the derived ``artists.csv``.
+        An artist keeps one row only, the last one of the file,
+        so that a re-fetched artist replaces its earlier row.  A
+        row whose name is not an artist of the store is dropped
+        and reported on the standard error.
 
-    :param file: The open snapshot CSV file, opened for append.
+        :param file: The open, seekable snapshot CSV file.
-    :param snapshot: The snapshot of an artist.
+        :param names: The artist names of the working store.
-    :return: None.
+        :return: None.
-    :raises OSError: When the file cannot be written.
+        :raises OSError: When the file cannot be read or written.
+        """
+        kept: dict[str, dict[str, str]] = {}
+        row: dict[str, str]
+        for row in ArtistSnapshotUpdater.__read_snapshot_rows(
+                file):
+            if row["name"] not in names:
+                print(f"dropped stale row \"{row['name']}\":"
+                      " no such artist in the store",
+                      file=sys.stderr)
+                continue
+            kept[row["name"]] = row
+        ordered: list[dict[str, str]] = sorted(
+            kept.values(), key=lambda x: x["name"].casefold())
+        file.seek(0)
+        file.truncate()
+        writer: csv.DictWriter[str] = csv.DictWriter(
+            file, ArtistSnapshotUpdater.__SNAPSHOT_FIELDS)
+        writer.writeheader()
+        writer.writerows(ordered)
+
+
+def parse_args(argv: list[str] | None) -> argparse.Namespace:
+    """Parse the command-line arguments.
+
+    :param argv: The command-line arguments, or None for
+        ``sys.argv``.
+    :return: The parsed arguments.
     """
-    csv.DictWriter(file, SNAPSHOT_FIELDS).writerow(
+    parser: argparse.ArgumentParser = argparse.ArgumentParser(
-        snapshot.to_row())
+        description="Fetch the artist metadata from Wikidata"
-    file.flush()
+                    " into the capture layer.")
-
+    parser.add_argument(
-
+        "wikidata_csv", type=Path,
-def write_snapshot(file: TextIO, names: Container[str]) -> None:
+        help="the Wikidata artist snapshot CSV file")
-    """Rewrite a snapshot CSV file handle sorted by artist name.
+    return parser.parse_args(argv)
-
-    The rows are ordered by the case-folded artist name, matching
-    the convention of the derived ``artists.csv``.  An artist
-    keeps one row only, the last one of the file, so that a
-    re-fetched artist replaces its earlier row.  A row whose name
-    is not an artist of the store is dropped and reported on the
-    standard error.
-
-    :param file: The open, seekable snapshot CSV file.
-    :param names: The artist names of the working store.
-    :return: None.
-    :raises OSError: When the file cannot be read or written.
-    """
-    kept: dict[str, dict[str, str]] = {}
-    row: dict[str, str]
-    for row in read_snapshot_rows(file):
-        if row["name"] not in names:
-            print(f"dropped stale row \"{row['name']}\":"
-                  " no such artist in the store", file=sys.stderr)
-            continue
-        kept[row["name"]] = row
-    ordered: list[dict[str, str]] = sorted(
-        kept.values(), key=lambda x: x["name"].casefold())
-    file.seek(0)
-    file.truncate()
-    writer: csv.DictWriter[str] = csv.DictWriter(
-        file, SNAPSHOT_FIELDS)
-    writer.writeheader()
-    writer.writerows(ordered)
 
 
 def main(argv: list[str] | None = None) -> int:
@@ -1066,52 +1179,16 @@ def main(argv: list[str] | None = None) -> int:
     """
     started: float = time.monotonic()
     args: argparse.Namespace = parse_args(argv)
-    fetcher: ArtistFetcher = ArtistFetcher()
-    fetched: int = 0
-    not_found: int = 0
-    errors: int = 0
-    session: Session = ds.get_db()
     try:
-        args.wikidata_csv.parent.mkdir(
+        counts: FetchCounts \
-            parents=True, exist_ok=True)
+            = ArtistSnapshotUpdater(args.wikidata_csv).run()
-        with open(args.wikidata_csv, "a+", encoding="utf-8",
-                  newline="") as csv_file:
-            done: set[str] = {x["name"] for x in
-                              read_snapshot_rows(csv_file)
-                              if x["gender"] != ""}
-            ensure_snapshot_header(csv_file)
-            names: set[str] = set()
-            artist: Artist
-            for artist in session.scalars(
-                    sa.select(Artist).order_by(Artist.id)):
-                names.add(artist.name)
-                if artist.name in done:
-                    continue
-                titles: list[str] = read_artist_titles(
-                    session, artist.id)
-                snapshot: ArtistSnapshot = fetcher.fetch(
-                    artist.name, titles)
-                append_row(csv_file, snapshot)
-                status: str = snapshot.qid
-                if snapshot.note == NOTE_NOT_FOUND:
-                    not_found += 1
-                    status = "not found"
-                elif snapshot.note.startswith("error: "):
-                    errors += 1
-                    status = snapshot.note
-                else:
-                    fetched += 1
-                print(f"artist \"{artist.name}\": {status}",
-                      file=sys.stderr)
-            write_snapshot(csv_file, names)
     except (OSError, sa.exc.SQLAlchemyError) as error:
         print(f"error: {error}", file=sys.stderr)
         return 1
-    finally:
+    attempted: int = counts.fetched + counts.not_found \
-        session.close()
+        + counts.errors
-    attempted: int = fetched + not_found + errors
     elapsed: str = format_duration(time.monotonic() - started)
-    print(f"Done.  Resolved {fetched}/{attempted} artists."
+    print(f"Done.  Resolved {counts.fetched}/{attempted}"
-          f"  {elapsed} elapsed.",
+          f" artists.  {elapsed} elapsed.",
           file=sys.stderr)
     return 0

tools/src/pop_fem_audit_tools/commands/fetch_lyrics.py

@@ -30,8 +30,9 @@ import time
 import urllib.parse
 import urllib.request
 from collections.abc import Sequence
+from dataclasses import dataclass
 from pathlib import Path
-from typing import Any
+from typing import Any, ClassVar
 
 import sqlalchemy as sa
 from sqlalchemy.orm import Session
@@ -45,60 +46,6 @@ from ..models import (
 )
 from ..utils import format_duration
 
-PROVENANCE_FIELDS: Sequence[str] = (
-    "song_id", "source", "method", "acquired_at", "note")
-"""The header columns of the lyrics provenance CSV file."""
-USER_AGENT: str = ("pop-fem-audit-tools"
-                   " (https://github.com/imacat/pop-fem-audit)")
-"""The User-Agent header sent on every HTTP request."""
-TIMEOUT: float = 30.0
-"""The timeout of an HTTP request, in seconds."""
-SLEEP_SECONDS: float = 1.0
-"""The delay between consecutive HTTP requests, in seconds."""
-
-
-def __build_normalization() -> dict[int, str | None]:
-    """Build the lyrics normalization translation table.
-
-    :return: The codepoint-to-replacement mapping, a replacement
-        of None meaning removal.
-    """
-    table: dict[int, str | None] = {}
-    codepoint: int
-    for codepoint in range(0x80, 0xa0):
-        try:
-            table[codepoint] = bytes([codepoint]).decode("cp1252")
-        except UnicodeDecodeError:
-            table[codepoint] = None
-    table[0x0435] = "e"
-    table[0x03cc] = "ó"
-    for codepoint in (0x2005, 0x205f, 0x200a):
-        table[codepoint] = " "
-    for codepoint in (0x200b, 0x200c, 0x200d, 0xfeff):
-        table[codepoint] = None
-    return table
-
-
-NORMALIZATION: dict[int, str | None] = __build_normalization()
-"""The codepoint-to-replacement mapping applied to fetched
-lyrics: cp1252-mojibake restoration for U+0080-U+009F (with the
-five byte values undefined in cp1252 removed), homoglyph
-restoration for the Cyrillic "e" and the Greek "o" with tonos,
-ASCII-space restoration for exotic space variants, and removal
-of zero-width characters.  A replacement of None removes the
-codepoint."""
-
-
-def normalize_lyrics(text: str) -> str:
-    """Restore or remove watermark and mojibake characters.
-
-    :param text: The lyrics text as fetched from an API.
-    :return: The text with the codepoints in
-        :data:`NORMALIZATION` replaced or removed; every other
-        character is unchanged.
-    """
-    return text.translate(NORMALIZATION)
-
 
 def parse_args(argv: list[str] | None) -> argparse.Namespace:
     """Parse the command-line arguments.
@@ -122,6 +69,16 @@ def parse_args(argv: list[str] | None) -> argparse.Namespace:
 class LyricsFetcher:
     """A fetcher of song lyrics from the public lyrics APIs."""
 
+    __USER_AGENT: ClassVar[str] = (
+        "pop-fem-audit-tools"
+        " (https://github.com/imacat/pop-fem-audit)")
+    """The User-Agent header sent on every HTTP request."""
+    __TIMEOUT: ClassVar[float] = 30.0
+    """The timeout of an HTTP request, in seconds."""
+    __SLEEP_SECONDS: ClassVar[float] = 1.0
+    """The delay between consecutive HTTP requests, in
+    seconds."""
+
     def __init__(self) -> None:
         """Construct the fetcher."""
         self.__sent: int = 0
@@ -193,78 +150,216 @@ class LyricsFetcher:
             network, or decoding error.
         """
         if self.__sent > 0:
-            time.sleep(SLEEP_SECONDS)
+            time.sleep(self.__SLEEP_SECONDS)
         self.__sent += 1
         request: urllib.request.Request = urllib.request.Request(
-            url, headers={"User-Agent": USER_AGENT})
+            url, headers={"User-Agent": self.__USER_AGENT})
         try:
             with urllib.request.urlopen(
-                    request, timeout=TIMEOUT) as response:
+                    request, timeout=self.__TIMEOUT) as response:
                 return json.load(response)
         except (OSError, ValueError):
             return None
 
 
-def query_artist(session: Session, song_id: int) -> str:
+@dataclass(frozen=True)
-    """Find the artist name to query the APIs with.
+class LyricsFetchCounts:
+    """The outcome of one run of fetching the missing lyrics."""
 
-    :param session: The database session.
+    fetched: int
-    :param song_id: The song ID.
+    """The number of songs newly fetched."""
-    :return: The name of the primary-role artist with the lowest
+    missed: int
-        position.
+    """The number of songs every API missed."""
-    """
-    name: str | None = session.scalar(
-        sa.select(Artist.name)
-        .join(SongArtist, SongArtist.artist_id == Artist.id)
-        .where(SongArtist.song_id == song_id,
-               SongArtist.role == Role.PRIMARY)
-        .order_by(SongArtist.position)
-        .limit(1))
-    assert name is not None
-    return name
 
 
-def save_lyrics(lyrics_dir: Path, song_id: int,
+class LyricsFetchRunner:
-                lyrics: str) -> None:
+    """The orchestrator of one run of fetching missing lyrics."""
-    """Write the lyrics of a song into the cache directory.
 
-    The cache directory is created when missing.
+    __PROVENANCE_FIELDS: ClassVar[Sequence[str]] = (
+        "song_id", "source", "method", "acquired_at", "note")
+    """The header columns of the lyrics provenance CSV file."""
 
-    The lyrics text is normalized with :func:`normalize_lyrics`
+    @staticmethod
-    before being written.
+    def __build_normalization() -> dict[int, str | None]:
+        """Build the lyrics normalization translation table.
 
-    :param lyrics_dir: The lyrics cache directory.
+        :return: The codepoint-to-replacement mapping, a
-    :param song_id: The song ID.
+            replacement of None meaning removal.
-    :param lyrics: The lyrics text.
+        """
-    :return: None.
+        table: dict[int, str | None] = {}
-    :raises OSError: When the file cannot be written.
+        codepoint: int
-    """
+        for codepoint in range(0x80, 0xa0):
-    lyrics_dir.mkdir(parents=True, exist_ok=True)
+            try:
-    (lyrics_dir / f"{song_id}.txt").write_text(
+                table[codepoint] = bytes(
-        normalize_lyrics(lyrics), encoding="utf-8")
+                    [codepoint]).decode("cp1252")
+            except UnicodeDecodeError:
+                table[codepoint] = None
+        table[0x0435] = "e"
+        table[0x03cc] = "ó"
+        for codepoint in (0x2005, 0x205f, 0x200a):
+            table[codepoint] = " "
+        for codepoint in (0x200b, 0x200c, 0x200d, 0xfeff):
+            table[codepoint] = None
+        return table
 
+    __NORMALIZATION: ClassVar[dict[int, str | None]] \
+        = __build_normalization()
+    """The codepoint-to-replacement mapping applied to fetched
+    lyrics: cp1252-mojibake restoration for U+0080-U+009F (with
+    the five byte values undefined in cp1252 removed), homoglyph
+    restoration for the Cyrillic "e" and the Greek "o" with
+    tonos, ASCII-space restoration for exotic space variants, and
+    removal of zero-width characters.  A replacement of None
+    removes the codepoint."""
 
-def append_provenance(path: Path, song_id: int,
+    def __init__(self, lyrics_dir: Path,
-                      source: str) -> None:
+                provenance_csv: Path) -> None:
-    """Append a provenance row for a fetched lyrics file.
+        """Set up the fetch run.
 
-    The CSV file is created with the header row when missing.
+        :param lyrics_dir: The lyrics cache directory.
+        :param provenance_csv: The lyrics provenance CSV file.
+        """
+        self.__lyrics_dir: Path = lyrics_dir
+        """The lyrics cache directory."""
+        self.__provenance_csv: Path = provenance_csv
+        """The lyrics provenance CSV file."""
+        self.__fetcher: LyricsFetcher = LyricsFetcher()
+        """The fetcher of the public lyrics APIs."""
 
-    :param path: The lyrics provenance CSV file.
+    def run(self) -> LyricsFetchCounts:
-    :param song_id: The song ID.
+        """Fetch the missing lyrics of every song in the store.
-    :param source: The source name of the fetched lyrics.
+
-    :return: None.
+        Every song fetched or missed is reported on the standard
-    :raises OSError: When the file cannot be written.
+        error as an observable side effect.
-    """
+
-    is_new: bool = not path.exists()
+        :return: The number of songs fetched and missed.
-    path.parent.mkdir(parents=True, exist_ok=True)
+        :raises OSError: When a cache file or the provenance CSV
-    with open(path, "a", encoding="utf-8",
+            cannot be written.
-              newline="") as file:
+        :raises sqlalchemy.exc.SQLAlchemyError: On a database
-        writer: Any = csv.writer(file)
+            error.
-        if is_new:
+        """
-            writer.writerow(PROVENANCE_FIELDS)
+        fetched: int = 0
-        writer.writerow([song_id, source, "api-fetch",
+        missed: int = 0
-                         datetime.date.today().isoformat(), ""])
+        session: Session = ds.get_db()
+        try:
+            song: Song
+            for song in session.scalars(
+                    sa.select(Song).order_by(Song.id)):
+                if (self.__lyrics_dir
+                        / f"{song.id}.txt").exists():
+                    continue
+                if self.__fetch_one(session, song):
+                    fetched += 1
+                else:
+                    missed += 1
+        finally:
+            session.close()
+        return LyricsFetchCounts(fetched=fetched, missed=missed)
+
+    def __fetch_one(self, session: Session, song: Song) -> bool:
+        """Fetch and save the lyrics of one song.
+
+        The song is queried by its primary-role artist name; when
+        every API misses and the song's full artist credit
+        differs from that name, the same APIs are queried again
+        with the artist credit.
+
+        :param session: The database session.
+        :param song: The song to fetch.
+        :return: True when a lyrics text was fetched and saved,
+            False when every API missed on both queries.
+        :raises OSError: When the cache file or the provenance
+            CSV cannot be written.
+        """
+        artist: str = self.__query_artist(session, song.id)
+        result: tuple[str, str] | None = self.__fetcher.fetch(
+            artist, song.title)
+        if result is None and song.artist_credit != artist:
+            result = self.__fetcher.fetch(
+                song.artist_credit, song.title)
+        if result is None:
+            print(f"song {song.id} \"{song.title}\": miss",
+                  file=sys.stderr)
+            return False
+        lyrics: str
+        source: str
+        lyrics, source = result
+        self.__save_lyrics(song.id, lyrics)
+        self.__append_provenance(song.id, source)
+        print(f"song {song.id} \"{song.title}\": {source}",
+              file=sys.stderr)
+        return True
+
+    @staticmethod
+    def __query_artist(session: Session, song_id: int) -> str:
+        """Find the artist name to query the APIs with.
+
+        :param session: The database session.
+        :param song_id: The song ID.
+        :return: The name of the primary-role artist with the
+            lowest position.
+        """
+        name: str | None = session.scalar(
+            sa.select(Artist.name)
+            .join(SongArtist, SongArtist.artist_id == Artist.id)
+            .where(SongArtist.song_id == song_id,
+                   SongArtist.role == Role.PRIMARY)
+            .order_by(SongArtist.position)
+            .limit(1))
+        assert name is not None
+        return name
+
+    def __save_lyrics(self, song_id: int, lyrics: str) -> None:
+        """Write the lyrics of a song into the cache directory.
+
+        The cache directory is created when missing.
+
+        The lyrics text is normalized with
+        :meth:`normalize_lyrics` before being written.
+
+        :param song_id: The song ID.
+        :param lyrics: The lyrics text.
+        :return: None.
+        :raises OSError: When the file cannot be written.
+        """
+        self.__lyrics_dir.mkdir(parents=True, exist_ok=True)
+        (self.__lyrics_dir / f"{song_id}.txt").write_text(
+            self.normalize_lyrics(lyrics), encoding="utf-8")
+
+    def __append_provenance(self, song_id: int,
+                            source: str) -> None:
+        """Append a provenance row for a fetched lyrics file.
+
+        The CSV file is created with the header row when
+        missing.
+
+        :param song_id: The song ID.
+        :param source: The source name of the fetched lyrics.
+        :return: None.
+        :raises OSError: When the file cannot be written.
+        """
+        is_new: bool = not self.__provenance_csv.exists()
+        self.__provenance_csv.parent.mkdir(
+            parents=True, exist_ok=True)
+        with open(self.__provenance_csv, "a", encoding="utf-8",
+                  newline="") as file:
+            writer: Any = csv.writer(file)
+            if is_new:
+                writer.writerow(self.__PROVENANCE_FIELDS)
+            writer.writerow(
+                [song_id, source, "api-fetch",
+                 datetime.date.today().isoformat(), ""])
+
+    @classmethod
+    def normalize_lyrics(cls, text: str) -> str:
+        """Restore or remove watermark and mojibake characters.
+
+        :param text: The lyrics text as fetched from an API.
+        :return: The text with the codepoints of the
+            normalization table replaced or removed; every other
+            character is unchanged.
+        """
+        return text.translate(cls.__NORMALIZATION)
 
 
 def main(argv: list[str] | None = None) -> int:
@@ -277,44 +372,15 @@ def main(argv: list[str] | None = None) -> int:
     """
     started: float = time.monotonic()
     args: argparse.Namespace = parse_args(argv)
-    fetcher: LyricsFetcher = LyricsFetcher()
-    fetched: int = 0
-    missed: int = 0
-    session: Session = ds.get_db()
     try:
-        song: Song
+        counts: LyricsFetchCounts = LyricsFetchRunner(
-        for song in session.scalars(
+            args.lyrics_dir, args.provenance_csv).run()
-                sa.select(Song).order_by(Song.id)):
-            if (args.lyrics_dir / f"{song.id}.txt").exists():
-                continue
-            artist: str = query_artist(session, song.id)
-            result: tuple[str, str] | None = fetcher.fetch(
-                artist, song.title)
-            if result is None and song.artist_credit != artist:
-                result = fetcher.fetch(
-                    song.artist_credit, song.title)
-            if result is None:
-                missed += 1
-                print(f"song {song.id} \"{song.title}\": miss",
-                      file=sys.stderr)
-                continue
-            lyrics: str
-            source: str
-            lyrics, source = result
-            save_lyrics(args.lyrics_dir, song.id, lyrics)
-            append_provenance(args.provenance_csv, song.id,
-                              source)
-            fetched += 1
-            print(f"song {song.id} \"{song.title}\": {source}",
-                  file=sys.stderr)
     except (OSError, sa.exc.SQLAlchemyError) as error:
         print(f"error: {error}", file=sys.stderr)
         return 1
-    finally:
+    attempted: int = counts.fetched + counts.missed
-        session.close()
-    attempted: int = fetched + missed
     elapsed: str = format_duration(time.monotonic() - started)
-    print(f"Done.  Fetched lyrics for {fetched}/{attempted}"
+    print(f"Done.  Fetched lyrics for {counts.fetched}/"
-          f" songs.  {elapsed} elapsed.",
+          f"{attempted} songs.  {elapsed} elapsed.",
           file=sys.stderr)
     return 0

tools/src/pop_fem_audit_tools/commands/run_llm.py

@@ -28,27 +28,13 @@ import time
 from dataclasses import asdict, dataclass
 from datetime import datetime
 from pathlib import Path
-from typing import Any, Self
+from typing import Any, ClassVar, Self
 
 import anthropic
 
 from ..config import get_settings
 from ..utils import format_duration
 
-# claude-fable-5 accepts neither "temperature" nor "thinking";
-# a model's entry holds exactly the extra request parameters it
-# accepts.
-MODELS: dict[str, dict[str, Any]] = {
-    "claude-sonnet-4-6": {
-        "temperature": 0.0,
-        "thinking": {"type": "disabled"},
-    },
-    "claude-fable-5": {},
-}
-DEFAULT_MODEL: str = "claude-sonnet-4-6"
-SCRIPT_VERSION: str = "run_llm.py 3.1.0"
-POLL_INTERVAL_SECONDS: float = 60.0
-
 
 class InputFormatError(Exception):
     """An error in the JSONL input file."""
@@ -136,13 +122,23 @@ class BatchResult:
                     if x.type == "text")
                 return cls(id=entry.custom_id, text=text,
                            stop_reason=message.stop_reason,
-                           usage=usage_to_dict(message.usage))
+                           usage=cls.__usage_to_dict(message.usage))
             case "errored":
                 return cls(id=entry.custom_id,
                            error=result.error.error.type)
             case other:
                 return cls(id=entry.custom_id, error=str(other))
 
+    @staticmethod
+    def __usage_to_dict(usage: Any) -> dict[str, Any]:
+        """Convert a usage object to a plain dictionary.
+
+        :param usage: The usage object of a message.
+        :return: The usage as a dictionary, without null entries.
+        """
+        return {k: v for k, v in usage.model_dump().items()
+                if v is not None}
+
     def to_record(self) -> dict[str, Any]:
         """Return this result as an archive JSONL record.
 
@@ -169,10 +165,409 @@ class BatchInfo:
     is still processing."""
 
 
+@dataclass(frozen=True)
+class ExecutionOutcome:
+    """The outcome of submitting and awaiting one batch."""
+
+    batch: BatchInfo
+    """The submitted batch's bookkeeping."""
+    results: Results
+    """The batch's results, keyed by item ID."""
+
+
+@dataclass(frozen=True)
+class RunOutcome:
+    """The outcome of one LLM definition file run."""
+
+    item_count: int
+    """The number of loaded input items."""
+    dry_run: bool
+    """Whether this was a dry run."""
+    dry_run_request: dict[str, Any] | None
+    """The first item's preview request, for a dry run; None for
+    an actual run."""
+    failed: list[str]
+    """The failed item IDs, in item order; always empty for a dry
+    run."""
+
+
+class LLMRunner:
+    """The orchestrator of one LLM definition file run."""
+
+    # claude-fable-5 accepts neither "temperature" nor "thinking";
+    # a model's entry holds exactly the extra request parameters
+    # it accepts.
+    MODELS: ClassVar[dict[str, dict[str, Any]]] = {
+        "claude-sonnet-4-6": {
+            "temperature": 0.0,
+            "thinking": {"type": "disabled"},
+        },
+        "claude-fable-5": {},
+    }
+    """The supported model IDs and their extra request
+    parameters."""
+    DEFAULT_MODEL: ClassVar[str] = "claude-sonnet-4-6"
+    """The default model ID."""
+    __SCRIPT_VERSION: ClassVar[str] = "run_llm.py 3.1.0"
+    """The script version recorded into the archive metadata."""
+    __POLL_INTERVAL_SECONDS: ClassVar[float] = 60.0
+    """The interval between batch status polls."""
+
+    def __init__(self, prompt: Path, input_path: Path,
+                 archive_dir: Path, model: str, max_tokens: int,
+                 dry_run: bool, replace: bool) -> None:
+        """Set up the run of one LLM definition file.
+
+        :param prompt: The prompt definition file, used as the
+            system prompt.
+        :param input_path: The JSONL input file with "id" and
+            "content".
+        :param archive_dir: The destination archive directory.
+        :param model: The model ID, a key of :attr:`MODELS`.
+        :param max_tokens: The maximum output tokens per request.
+        :param dry_run: Whether to validate and archive without
+            calling the API.
+        :param replace: Whether to replace an already existing
+            archive directory.
+        """
+        self.__prompt: Path = prompt
+        """The prompt definition file."""
+        self.__input: Path = input_path
+        """The JSONL input file."""
+        self.__archive_dir: Path = archive_dir
+        """The destination archive directory."""
+        self.__model: str = model
+        """The model ID."""
+        self.__max_tokens: int = max_tokens
+        """The maximum output tokens per request."""
+        self.__dry_run: bool = dry_run
+        """Whether to validate and archive without calling the
+        API."""
+        self.__replace: bool = replace
+        """Whether to replace an already existing archive
+        directory."""
+
+    def run(self) -> RunOutcome:
+        """Load the input, archive the prompt, and run the batch.
+
+        Always writes ``prompt.md`` and ``meta.json`` into the
+        archive directory.  A dry run stops there, previewing the
+        first item's request; an actual run also submits the
+        batch, awaits it, and writes ``output.jsonl``.
+
+        :return: The outcome of the run.
+        :raises InputFormatError: When the input file is
+            malformed.
+        :raises OSError: When the input or prompt file cannot be
+            read, the archive directory already exists without
+            ``replace``, or an output file cannot be written.
+        """
+        items: list[InputItem] = self.__load_items()
+        prompt_text: str = self.__prompt.read_text(encoding="utf-8")
+        archive_dir: Path = self.__create_archive_dir()
+        (archive_dir / "prompt.md").write_bytes(
+            self.__prompt.read_bytes())
+        meta: dict[str, Any] = self.__build_meta(items)
+        meta_path: Path = archive_dir / "meta.json"
+        if self.__dry_run:
+            self.__write_json(meta_path, meta)
+            request: dict[str, Any] = self.__build_request(
+                items[0], prompt_text)
+            return RunOutcome(
+                item_count=len(items), dry_run=True,
+                dry_run_request=request, failed=[])
+        client: anthropic.Anthropic = anthropic.Anthropic(
+            api_key=get_settings().ANTHROPIC_API_KEY)
+        outcome: ExecutionOutcome = self.__execute_run(
+            client, items, prompt_text)
+        item_ids: list[str] = [x.id for x in items]
+        self.__write_jsonl(
+            archive_dir / "output.jsonl",
+            [outcome.results[x].to_record() for x in item_ids
+             if x in outcome.results])
+        meta["batch"] = outcome.batch
+        meta["usage"] = self.__sum_usage(outcome.results)
+        self.__write_meta(meta_path, meta)
+        failed: list[str] = self.__find_failures(
+            item_ids, outcome.results)
+        return RunOutcome(
+            item_count=len(items), dry_run=False,
+            dry_run_request=None, failed=failed)
+
+    def __load_items(self) -> list[InputItem]:
+        """Load and validate the JSONL input items.
+
+        :return: The input items, in file order.
+        :raises InputFormatError: When a line is malformed, an ID
+            is duplicated, or the file contains no item.
+        :raises OSError: When the file cannot be read.
+        """
+        items: list[InputItem] = []
+        seen: set[str] = set()
+        with open(self.__input, encoding="utf-8") as file:
+            for number, line in enumerate(file, start=1):
+                if line.strip() == "":
+                    continue
+                data: Any
+                try:
+                    data = json.loads(line)
+                except json.JSONDecodeError as error:
+                    raise InputFormatError(
+                        f"{self.__input}: line {number}: malformed"
+                        f" JSON: {error}")
+                item: InputItem = InputItem.get_instance(
+                    data, self.__input, number)
+                if item.id in seen:
+                    raise InputFormatError(
+                        f"{self.__input}: line {number}:"
+                        f" duplicated ID \"{item.id}\"")
+                seen.add(item.id)
+                items.append(item)
+        if len(items) == 0:
+            raise InputFormatError(f"{self.__input}: no input items")
+        return items
+
+    def __create_archive_dir(self) -> Path:
+        """Create the archive directory.
+
+        Only this directory is ever created or removed; no other
+        directory is ever touched.
+
+        :return: The created archive directory.
+        :raises FileExistsError: When the archive directory
+            already exists and ``replace`` is False.
+        """
+        if self.__archive_dir.exists():
+            if not self.__replace:
+                raise FileExistsError(
+                    f"{self.__archive_dir} already exists; pass"
+                    " --replace to replace it")
+            shutil.rmtree(self.__archive_dir)
+        self.__archive_dir.mkdir(parents=True)
+        return self.__archive_dir
+
+    def __build_meta(self, items: list[InputItem]) -> dict[str, Any]:
+        """Build the initial archive metadata.
+
+        :param items: The loaded input items.
+        :return: The metadata, "batch" and "usage" not yet filled
+            in for an actual run.
+        """
+        return {
+            "script_version": self.__SCRIPT_VERSION,
+            "model": self.__model,
+            "temperature": self.MODELS[self.__model].get(
+                "temperature"),
+            "thinking": self.MODELS[self.__model].get("thinking"),
+            "max_tokens": self.__max_tokens,
+            "prompt_path": str(self.__prompt),
+            "prompt_sha256": self.__sha256_of(self.__prompt),
+            "input_path": str(self.__input),
+            "input_sha256": self.__sha256_of(self.__input),
+            "item_count": len(items),
+            "dry_run": self.__dry_run,
+            "started_at": self.__now_iso(),
+            "batch": None,
+            "usage": {},
+        }
+
+    def __build_request(self, item: InputItem, system_prompt: str) \
+            -> dict[str, Any]:
+        """Build one Message Batches request for an input item.
+
+        :param item: The input item.
+        :param system_prompt: The system prompt text.
+        :return: The batch request with "custom_id" and "params".
+        """
+        return {
+            "custom_id": item.id,
+            "params": {
+                "model": self.__model,
+                "max_tokens": self.__max_tokens,
+                **self.MODELS[self.__model],
+                "system": system_prompt,
+                "messages": [
+                    {"role": "user", "content": item.content},
+                ],
+            },
+        }
+
+    def __execute_run(
+            self, client: anthropic.Anthropic,
+            items: list[InputItem], system_prompt: str) \
+            -> ExecutionOutcome:
+        """Submit the batch of this run and await its results.
+
+        :param client: The Anthropic client.
+        :param items: The input items.
+        :param system_prompt: The system prompt text.
+        :return: The submitted batch's bookkeeping and its
+            results.
+        """
+        requests: list[dict[str, Any]] = [
+            self.__build_request(x, system_prompt) for x in items]
+        info: BatchInfo = BatchInfo(
+            batch_id=self.__submit_batch(client, requests),
+            submitted_at=self.__now_iso())
+        print(f"submitted batch {info.batch_id}", file=sys.stderr)
+        batches: dict[str, Any] = self.__poll_batches(
+            client, [info.batch_id])
+        info.ended_at = batches[info.batch_id].ended_at.isoformat()
+        results: Results = self.__collect_results(
+            client, info.batch_id)
+        return ExecutionOutcome(batch=info, results=results)
+
+    @staticmethod
+    def __submit_batch(client: anthropic.Anthropic,
+                       requests: list[dict[str, Any]]) -> str:
+        """Submit one message batch.
+
+        :param client: The Anthropic client.
+        :param requests: The batch requests.
+        :return: The batch ID.
+        """
+        return client.messages.batches.create(requests=requests).id
+
+    @classmethod
+    def __poll_batches(cls, client: anthropic.Anthropic,
+                       batch_ids: list[str]) -> dict[str, Any]:
+        """Poll the batches until every one of them has ended.
+
+        Progress is printed to the standard error every poll.
+
+        :param client: The Anthropic client.
+        :param batch_ids: The batch IDs to poll.
+        :return: The final batch object of each batch, keyed by
+            batch ID.
+        """
+        while True:
+            batches: dict[str, Any] = {
+                x: client.messages.batches.retrieve(x)
+                for x in batch_ids}
+            pending: list[str] = [
+                x for x in batch_ids
+                if batches[x].processing_status != "ended"]
+            for batch_id in batch_ids:
+                status: str = batches[batch_id].processing_status
+                print(f"batch {batch_id}: {status}", file=sys.stderr)
+            if len(pending) == 0:
+                return batches
+            time.sleep(cls.__POLL_INTERVAL_SECONDS)
+
+    @staticmethod
+    def __collect_results(client: anthropic.Anthropic,
+                          batch_id: str) -> Results:
+        """Collect the results of an ended batch.
+
+        :param client: The Anthropic client.
+        :param batch_id: The batch ID.
+        :return: The result records, keyed by custom ID.
+        """
+        results: Results = {}
+        for entry in client.messages.batches.results(batch_id):
+            results[entry.custom_id] = BatchResult.get_instance(
+                entry)
+        return results
+
+    @staticmethod
+    def __find_failures(item_ids: list[str],
+                        results: Results) -> list[str]:
+        """Find the item IDs that failed in a result set.
+
+        An item failed when it is missing from the results or when
+        its record is a failure.
+
+        :param item_ids: The item IDs to check, in order.
+        :param results: The result records, keyed by item ID.
+        :return: The failed item IDs, in the given order.
+        """
+        return [x for x in item_ids
+                if x not in results or results[x].is_failure]
+
+    @staticmethod
+    def __sum_usage(results: Results) -> dict[str, int]:
+        """Sum the token usage of every succeeded result.
+
+        :param results: The result records, keyed by item ID.
+        :return: The summed integer usage fields.
+        """
+        totals: dict[str, int] = {}
+        for result in results.values():
+            if result.usage is None:
+                continue
+            for key, value in result.usage.items():
+                if isinstance(value, int):
+                    totals[key] = totals.get(key, 0) + value
+        return totals
+
+    @staticmethod
+    def __write_jsonl(path: Path,
+                      records: list[dict[str, Any]]) -> None:
+        """Write records to a file as JSON Lines.
+
+        :param path: The path of the file to write.
+        :param records: The records, one per line.
+        :return: None.
+        :raises OSError: When the file cannot be written.
+        """
+        with open(path, "w", encoding="utf-8") as file:
+            for record in records:
+                file.write(
+                    json.dumps(record, ensure_ascii=False) + "\n")
+
+    @staticmethod
+    def __write_json(path: Path, data: dict[str, Any]) -> None:
+        """Write data to a file as pretty-printed JSON.
+
+        :param path: The path of the file to write.
+        :param data: The data to write.
+        :return: None.
+        :raises OSError: When the file cannot be written.
+        """
+        path.write_text(
+            json.dumps(data, ensure_ascii=False, indent=2) + "\n",
+            encoding="utf-8")
+
+    @classmethod
+    def __write_meta(cls, path: Path, meta: dict[str, Any]) -> None:
+        """Write the metadata to the ``meta.json`` file.
+
+        The ``BatchInfo`` value under "batch" is written as a
+        plain JSON object.
+
+        :param path: The path of the ``meta.json`` file.
+        :param meta: The metadata to write.
+        :return: None.
+        :raises OSError: When the file cannot be written.
+        """
+        cls.__write_json(
+            path, {**meta, "batch": asdict(meta["batch"])})
+
+    @staticmethod
+    def __sha256_of(path: Path) -> str:
+        """Calculate the SHA-256 digest of a file.
+
+        :param path: The path of the file.
+        :return: The hexadecimal SHA-256 digest.
+        """
+        with open(path, "rb") as file:
+            return hashlib.file_digest(file, "sha256").hexdigest()
+
+    @staticmethod
+    def __now_iso() -> str:
+        """Return the current local time in ISO 8601 format.
+
+        :return: The current local time with the timezone offset.
+        """
+        return datetime.now().astimezone().isoformat(
+            timespec="seconds")
+
+
 def parse_args(argv: list[str] | None) -> argparse.Namespace:
     """Parse the command-line arguments.
 
-    :param argv: The command-line arguments, or None for ``sys.argv``.
+    :param argv: The command-line arguments, or None for
+        ``sys.argv``.
     :return: The parsed arguments.
     """
     parser: argparse.ArgumentParser = argparse.ArgumentParser(
@@ -180,7 +575,8 @@ def parse_args(argv: list[str] | None) -> argparse.Namespace:
                     " and archive the result.")
     parser.add_argument(
         "prompt", type=Path,
-        help="the prompt definition file, used as the system prompt")
+        help="the prompt definition file, used as the system"
+             " prompt")
     parser.add_argument(
         "input", type=Path,
         help="the JSONL input file with \"id\" and \"content\"")
@@ -188,11 +584,13 @@ def parse_args(argv: list[str] | None) -> argparse.Namespace:
         "archive_dir", type=Path,
         help="the destination archive directory")
     parser.add_argument(
-        "--model", choices=sorted(MODELS), default=DEFAULT_MODEL,
+        "--model", choices=sorted(LLMRunner.MODELS),
-        help=f"the model ID (default {DEFAULT_MODEL})")
+        default=LLMRunner.DEFAULT_MODEL,
+        help=f"the model ID (default {LLMRunner.DEFAULT_MODEL})")
     parser.add_argument(
         "--max-tokens", type=int, default=2048,
-        help="the maximum output tokens per request (default 2048)")
+        help="the maximum output tokens per request (default"
+             " 2048)")
     parser.add_argument(
         "--dry-run", action="store_true",
         help="validate and archive without calling the API")
@@ -202,327 +600,30 @@ def parse_args(argv: list[str] | None) -> argparse.Namespace:
     return parser.parse_args(argv)
 
 
-def load_items(path: Path) -> list[InputItem]:
-    """Load and validate the JSONL input items.
-
-    :param path: The path of the JSONL input file.
-    :return: The input items, in file order.
-    :raises InputFormatError: When a line is malformed, an ID is
-        duplicated, or the file contains no item.
-    :raises OSError: When the file cannot be read.
-    """
-    items: list[InputItem] = []
-    seen: set[str] = set()
-    with open(path, encoding="utf-8") as file:
-        for number, line in enumerate(file, start=1):
-            if line.strip() == "":
-                continue
-            try:
-                data: Any = json.loads(line)
-            except json.JSONDecodeError as error:
-                raise InputFormatError(
-                    f"{path}: line {number}: malformed JSON: {error}")
-            item: InputItem = InputItem.get_instance(
-                data, path, number)
-            if item.id in seen:
-                raise InputFormatError(
-                    f"{path}: line {number}: duplicated ID"
-                    f" \"{item.id}\"")
-            seen.add(item.id)
-            items.append(item)
-    if len(items) == 0:
-        raise InputFormatError(f"{path}: no input items")
-    return items
-
-
-def build_request(item: InputItem, system_prompt: str,
-                  max_tokens: int, model: str) -> dict[str, Any]:
-    """Build one Message Batches request for an input item.
-
-    :param item: The input item.
-    :param system_prompt: The system prompt text.
-    :param max_tokens: The maximum output tokens.
-    :param model: The model ID, a key of ``MODELS``.
-    :return: The batch request with "custom_id" and "params".
-    """
-    return {
-        "custom_id": item.id,
-        "params": {
-            "model": model,
-            "max_tokens": max_tokens,
-            **MODELS[model],
-            "system": system_prompt,
-            "messages": [
-                {"role": "user", "content": item.content},
-            ],
-        },
-    }
-
-
-def submit_batch(client: anthropic.Anthropic,
-                 requests: list[dict[str, Any]]) -> str:
-    """Submit one message batch.
-
-    :param client: The Anthropic client.
-    :param requests: The batch requests.
-    :return: The batch ID.
-    """
-    return client.messages.batches.create(requests=requests).id
-
-
-def poll_batches(client: anthropic.Anthropic,
-                 batch_ids: list[str]) -> dict[str, Any]:
-    """Poll the batches until every one of them has ended.
-
-    Progress is printed to the standard error every poll.
-
-    :param client: The Anthropic client.
-    :param batch_ids: The batch IDs to poll.
-    :return: The final batch object of each batch, keyed by batch ID.
-    """
-    while True:
-        batches: dict[str, Any] = {
-            x: client.messages.batches.retrieve(x) for x in batch_ids}
-        pending: list[str] = [
-            x for x in batch_ids
-            if batches[x].processing_status != "ended"]
-        for batch_id in batch_ids:
-            status: str = batches[batch_id].processing_status
-            print(f"batch {batch_id}: {status}", file=sys.stderr)
-        if len(pending) == 0:
-            return batches
-        time.sleep(POLL_INTERVAL_SECONDS)
-
-
-def usage_to_dict(usage: Any) -> dict[str, Any]:
-    """Convert a usage object to a plain dictionary.
-
-    :param usage: The usage object of a message.
-    :return: The usage as a dictionary, without null entries.
-    """
-    return {k: v for k, v in usage.model_dump().items()
-            if v is not None}
-
-
-def sum_usage(results: Results) -> dict[str, int]:
-    """Sum the token usage of every succeeded result.
-
-    :param results: The result records, keyed by item ID.
-    :return: The summed integer usage fields.
-    """
-    totals: dict[str, int] = {}
-    for result in results.values():
-        if result.usage is None:
-            continue
-        for key, value in result.usage.items():
-            if isinstance(value, int):
-                totals[key] = totals.get(key, 0) + value
-    return totals
-
-
-def collect_results(client: anthropic.Anthropic,
-                    batch_id: str) -> Results:
-    """Collect the results of an ended batch.
-
-    :param client: The Anthropic client.
-    :param batch_id: The batch ID.
-    :return: The result records, keyed by custom ID.
-    """
-    results: Results = {}
-    for entry in client.messages.batches.results(batch_id):
-        results[entry.custom_id] = BatchResult.get_instance(entry)
-    return results
-
-
-def find_failures(item_ids: list[str],
-                  results: Results) -> list[str]:
-    """Find the item IDs that failed in a result set.
-
-    An item failed when it is missing from the results or when its
-    record is a failure.
-
-    :param item_ids: The item IDs to check, in order.
-    :param results: The result records, keyed by item ID.
-    :return: The failed item IDs, in the given order.
-    """
-    return [x for x in item_ids
-            if x not in results or results[x].is_failure]
-
-
-def create_archive_dir(directory: Path, replace: bool) -> Path:
-    """Create the archive directory.
-
-    Only this directory is ever created or removed; no other
-    directory is ever touched.
-
-    :param directory: The destination archive directory.
-    :param replace: Whether to remove an already existing archive
-        directory before creating it.
-    :return: The created archive directory.
-    :raises FileExistsError: When the archive directory already
-        exists and ``replace`` is False.
-    """
-    if directory.exists():
-        if not replace:
-            raise FileExistsError(
-                f"{directory} already exists; pass --replace to"
-                " replace it")
-        shutil.rmtree(directory)
-    directory.mkdir(parents=True)
-    return directory
-
-
-def write_jsonl(path: Path, records: list[dict[str, Any]]) -> None:
-    """Write records to a file as JSON Lines.
-
-    :param path: The path of the file to write.
-    :param records: The records, one per line.
-    :return: None.
-    """
-    with open(path, "w", encoding="utf-8") as file:
-        for record in records:
-            file.write(json.dumps(record, ensure_ascii=False) + "\n")
-
-
-def write_json(path: Path, data: dict[str, Any]) -> None:
-    """Write data to a file as pretty-printed JSON.
-
-    :param path: The path of the file to write.
-    :param data: The data to write.
-    :return: None.
-    """
-    path.write_text(
-        json.dumps(data, ensure_ascii=False, indent=2) + "\n",
-        encoding="utf-8")
-
-
-def write_meta(path: Path, meta: dict[str, Any]) -> None:
-    """Write the metadata to the ``meta.json`` file.
-
-    The ``BatchInfo`` value under ``batch`` is written as a plain
-    JSON object.
-
-    :param path: The path of the ``meta.json`` file.
-    :param meta: The metadata to write.
-    :return: None.
-    """
-    write_json(path, {**meta, "batch": asdict(meta["batch"])})
-
-
-def sha256_of(path: Path) -> str:
-    """Calculate the SHA-256 digest of a file.
-
-    :param path: The path of the file.
-    :return: The hexadecimal SHA-256 digest.
-    """
-    with open(path, "rb") as file:
-        return hashlib.file_digest(file, "sha256").hexdigest()
-
-
-def now_iso() -> str:
-    """Return the current local time in ISO 8601 format.
-
-    :return: The current local time with the timezone offset.
-    """
-    return datetime.now().astimezone().isoformat(timespec="seconds")
-
-
-def execute_run(
-        client: anthropic.Anthropic, items: list[InputItem],
-        system_prompt: str, max_tokens: int, model: str,
-        meta: dict[str, Any],
-) -> Results:
-    """Submit the batch of this run and await its results.
-
-    The batch ID and timestamps are recorded into the metadata as an
-    observable side effect.
-
-    :param client: The Anthropic client.
-    :param items: The input items.
-    :param system_prompt: The system prompt text.
-    :param max_tokens: The maximum output tokens per request.
-    :param model: The model ID, a key of ``MODELS``.
-    :param meta: The metadata to record the batch bookkeeping into.
-    :return: The results of this run, keyed by item ID.
-    """
-    requests: list[dict[str, Any]] = [
-        build_request(x, system_prompt, max_tokens, model)
-        for x in items]
-    info: BatchInfo = BatchInfo(
-        batch_id=submit_batch(client, requests),
-        submitted_at=now_iso())
-    meta["batch"] = info
-    print(f"submitted batch {info.batch_id}", file=sys.stderr)
-    batches: dict[str, Any] = poll_batches(client, [info.batch_id])
-    info.ended_at = batches[info.batch_id].ended_at.isoformat()
-    return collect_results(client, info.batch_id)
-
-
 def main(argv: list[str] | None = None) -> int:
     """Run one LLM definition file against one input and archive it.
 
-    :param argv: The command-line arguments, or None for ``sys.argv``.
+    :param argv: The command-line arguments, or None for
+        ``sys.argv``.
     :return: The exit status: 0 on success, non-zero on failure.
     """
     started: float = time.monotonic()
     args: argparse.Namespace = parse_args(argv)
     try:
-        items: list[InputItem] = load_items(args.input)
+        outcome: RunOutcome = LLMRunner(
-        prompt_text: str = args.prompt.read_text(encoding="utf-8")
+            args.prompt, args.input, args.archive_dir, args.model,
-    except (OSError, InputFormatError) as error:
+            args.max_tokens, args.dry_run, args.replace).run()
+    except (InputFormatError, OSError) as error:
         print(f"error: {error}", file=sys.stderr)
         return 1
-    try:
+    if not outcome.dry_run and len(outcome.failed) > 0:
-        archive_dir: Path = create_archive_dir(
+        print(f"error: failed items: {', '.join(outcome.failed)}",
-            args.archive_dir, args.replace)
-    except FileExistsError as error:
-        print(f"error: {error}", file=sys.stderr)
-        return 1
-    meta_path: Path = archive_dir / "meta.json"
-    (archive_dir / "prompt.md").write_bytes(args.prompt.read_bytes())
-    meta: dict[str, Any] = {
-        "script_version": SCRIPT_VERSION,
-        "model": args.model,
-        "temperature": MODELS[args.model].get("temperature"),
-        "thinking": MODELS[args.model].get("thinking"),
-        "max_tokens": args.max_tokens,
-        "prompt_path": str(args.prompt),
-        "prompt_sha256": sha256_of(args.prompt),
-        "input_path": str(args.input),
-        "input_sha256": sha256_of(args.input),
-        "item_count": len(items),
-        "dry_run": args.dry_run,
-        "started_at": now_iso(),
-        "batch": None,
-        "usage": {},
-    }
-    if args.dry_run:
-        write_json(meta_path, meta)
-        print(json.dumps(
-            build_request(items[0], prompt_text, args.max_tokens,
-                          args.model),
-            ensure_ascii=False, indent=2))
-        elapsed: str = format_duration(time.monotonic() - started)
-        print(f"Done.  {len(items)} jobs finished."
-              f"  {elapsed} elapsed.", file=sys.stderr)
-        return 0
-    client: anthropic.Anthropic = anthropic.Anthropic(
-        api_key=get_settings().ANTHROPIC_API_KEY)
-    results: Results = execute_run(
-        client, items, prompt_text, args.max_tokens, args.model,
-        meta)
-    item_ids: list[str] = [x.id for x in items]
-    write_jsonl(
-        archive_dir / "output.jsonl",
-        [results[x].to_record() for x in item_ids if x in results])
-    meta["usage"] = sum_usage(results)
-    write_meta(meta_path, meta)
-    failed: list[str] = find_failures(item_ids, results)
-    if len(failed) > 0:
-        print(f"error: failed items: {', '.join(failed)}",
               file=sys.stderr)
         return 1
+    if outcome.dry_run:
+        print(json.dumps(
+            outcome.dry_run_request, ensure_ascii=False, indent=2))
     elapsed: str = format_duration(time.monotonic() - started)
-    print(f"Done.  {len(items)} jobs finished."
+    print(f"Done.  {outcome.item_count} jobs finished."
           f"  {elapsed} elapsed.", file=sys.stderr)
     return 0