Add step 4 semantic code grouping by majority vote on claude-fable-5

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-08-17 22:38:40 +08:00
parent 683b15e073
commit c3d6c06910
14 changed files with 294 additions and 17 deletions
@@ -3,8 +3,11 @@
 ## Analysis pipeline

 - LLM analysis runs via Python scripts calling the Anthropic
-  Messages API: model `claude-sonnet-4-6`, `temperature=0`,
-  thinking disabled, Batch API where possible.
+  Messages API, Batch API where possible.  Steps 1 and 3 run
+  on `claude-sonnet-4-6` with `temperature=0` and thinking
+  disabled; step 4 runs on `claude-fable-5`, which accepts
+  neither parameter -- its sampling variance is absorbed by
+  the majority vote.
 - 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
@@ -13,11 +16,13 @@
  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".
- Per-song LLM judgments (coding) run the same definition
-  file three times, independently, over the same input; a
-  deterministic subcommand then assigns a (song, keyword)
-  pair when at least two of the three runs assign it
-  ("3 runs + majority vote").  Free-generation steps run
+- 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 vocabulary is
  built by a deterministic subcommand (embedding +
  clustering), not by an LLM.  If a validation outcome is
@@ -688,3 +688,41 @@
  檢定(women-power/female-empowerment 對三群之 2×2,
  Fisher 精確檢定,6 檢定一族之 BH-FDR 校正 q)之數字
  以本條所載分群為據,論文之編碼群表與群關聯表同此。
+
+- **步驟 4:語意編碼群以三票制定案(claude-fable-5),
+  取代草稿分群**:更正前條——所稱「研究者審定」實未
+  發生,研究者僅將盲選輸出整理歸檔,未作實質裁決;草稿
+  分群(單次 Claude Code 盲選)不具方法學地位。裁定
+  步驟 4 為語意分群的正式程序。任務與盲選同構:每筆輸入
+  為群名加 101 碼字母序清單,輸出為入選碼的單層 JSON
+  陣列;定義檔 `prompts/04-group.md` 只定格式、不含群的
+  語意定義。模型裁定 claude-fable-5:模型對照實驗(同
+  定義檔、同輸入、Batch API)顯示分群對模型高度敏感——
+  claude-sonnet-4-6 將群名成份式拆讀(women+power,凡
+  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 對照執行歸檔另行私人備份,不入
+  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
+  無效。三次執行(`runs/04-group/run1`–`run3`)多數決
+  定案(22 個(群,碼)中 21 個三次全票,
+  family-and-fatherhood 以 2:1 入群):女性力量群 2 碼
+  ——female-empowerment、women-power;厭女群 1 碼——
+  rejection-of-women;陽剛男性氣質群 9 碼——
+  dominance-and-power、family-and-fatherhood、
+  hustle-and-money、rivalry-and-superiority、
+  self-confidence-and-braggadocio、
+  showing-off-and-impressing、street-loyalty-and-danger、
+  violence-and-street-danger、wealth-and-flexing;脆弱群
+  10 碼——disappointment-and-failure、
+  fear-of-losing-love、heartbreak-and-grief、
+  hidden-emotional-struggle、inner-mental-turmoil、
+  loneliness-and-isolation、longing-and-loss、
+  past-trauma-and-healing、self-worth-and-insecurity、
+  vulnerability-and-betrayal。計票由確定性子命令
+  `tally-groups` 重現,定案寫入 `results/groups.csv`
+  (群、編碼、票數);論文之編碼群表、上標註記與群層次
+  檢定改以本定案為據,相關數表隨之重算。
@@ -6,17 +6,18 @@

 ## 自然編碼管線總覽

-三個步驟：步驟 1 自由標註（兩次執行進池）→ 步驟 2 詞彙表
+四個步驟：步驟 1 自由標註（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）。歌詞只出現在步驟 1 與步驟 3；步驟 2 完全不接觸
-歌詞，也不呼叫 LLM。設計原則見 `research-plan.md`；本檔
+多數決）→ 步驟 4 語意編碼群（三次執行＋多數決）。歌詞只
+出現在步驟 1 與步驟 3；步驟 2 與步驟 4 完全不接觸歌詞，
+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 與
-步驟 3 有定義檔（`prompts/`），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。
+編號的所指為**研究程序的工序**，不是定義檔：步驟 1、
+步驟 3 與步驟 4 有定義檔（`prompts/`），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

@@ -113,6 +114,27 @@
  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
+
+- **對象**：將 101 個編碼依語意劃入研究者命名的四個編碼
+  群——女性力量（women-power group）、反女性力量／厭女
+  （misogyny group）、陽剛男性氣質（masculine group）、
+  脆弱（vulnerable group）。群只有名字，沒有定義；歸屬
+  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。
+- **模型**：`claude-fable-5`（步驟 1、3 為
+  `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`）。
+- **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。計票由
+  確定性子命令 `tally-groups` 完成，定案分群寫入
+  `results/groups.csv`（欄位：群、編碼、票數）。
+
 ## 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，
@@ -188,7 +210,7 @@

 - LLM 步驟以 `run-llm <定義檔> <輸入檔> <歸檔目錄>`
  執行；一步的 N 次執行＝重現命令清單上的 N 行命令，
-  各自歸檔（`runs/<步驟>/run1`、`run2`，編碼步驟另有
+  各自歸檔（`runs/<步驟>/run1`、`run2`，三票制步驟另有
  `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
@@ -39,5 +39,10 @@ $3／$15、opus-4-6 $5／$25、opus-5 與 fable-5 $10／$50）
 | 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；歸檔另行私人備份，不入版本庫） |
+| 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） |

-累計支出：$63.22。
+累計支出：$63.48。
@@ -0,0 +1,30 @@
+You are selecting thematic keywords that belong to a named
+group.
+
+Input: a JSON object with the name of one group and the
+keywords — the complete set of codes; use these and no
+others:
+
+{
+  "group": "the name of the group",
+  "keywords": ["first-keyword", "second-keyword"]
+}
+
+Task: list every given keyword that belongs to the named
+group, judged by the literal meaning of the keyword itself.
+
+Rules:
+
+- Use only the given keywords, spelled exactly as given.
+- A group may match any number of keywords, including none.
+- Judge each keyword only by the literal meaning of its own
+  wording.
+
+Do not wrap the output in a Markdown code fence.
+The output must be strictly valid JSON.
+
+Output a single JSON array of the keywords that belong to
+the group (an empty array when none belongs), and nothing
+else:
+
+["first-keyword", "second-keyword"]
@@ -0,0 +1,25 @@
+{
+  "script_version": "run_llm.py 3.0.0",
+  "model": "claude-fable-5",
+  "temperature": null,
+  "thinking": null,
+  "max_tokens": 8192,
+  "prompt_path": "../prompts/04-group.md",
+  "prompt_sha256": "50e1d5d6c4115d090ae4ab1e801ea74db0b2dcdd47213ce0feb6749e034a41e9",
+  "input_path": "instance/llm-input-group.jsonl",
+  "input_sha256": "7fd97ae165270cb91461aedf411050892667b928d8584fceffea6f9649e003c7",
+  "item_count": 4,
+  "dry_run": false,
+  "started_at": "2026-08-14T22:30:32+08:00",
+  "batch": {
+    "batch_id": "msgbatch_01Mr6goBb2Efa4YrqCprbP4U",
+    "submitted_at": "2026-08-14T22:30:34+08:00",
+    "ended_at": "2026-08-14T14:31:27.547545+00:00"
+  },
+  "usage": {
+    "cache_creation_input_tokens": 0,
+    "cache_read_input_tokens": 0,
+    "input_tokens": 5553,
+    "output_tokens": 2049
+  }
+}
@@ -0,0 +1,4 @@
+{"id": "group-women-power", "text": "[\"female-empowerment\", \"women-power\"]", "stop_reason": "end_turn", "usage": {"cache_creation": {"ephemeral_1h_input_tokens": 0, "ephemeral_5m_input_tokens": 0}, "cache_creation_input_tokens": 0, "cache_read_input_tokens": 0, "inference_geo": "global", "input_tokens": 1385, "output_tokens": 53, "output_tokens_details": {"thinking_tokens": 37}, "service_tier": "batch"}}
+{"id": "group-misogyny", "text": "[\"rejection-of-women\"]", "stop_reason": "end_turn", "usage": {"cache_creation": {"ephemeral_1h_input_tokens": 0, "ephemeral_5m_input_tokens": 0}, "cache_creation_input_tokens": 0, "cache_read_input_tokens": 0, "inference_geo": "global", "input_tokens": 1391, "output_tokens": 118, "output_tokens_details": {"thinking_tokens": 106}, "service_tier": "batch"}}
+{"id": "group-masculine", "text": "[\"avoiding-commitment\", \"dominance-and-power\", \"hustle-and-money\", \"rejection-of-women\", \"rivalry-and-superiority\", \"self-confidence-and-braggadocio\", \"showing-off-and-impressing\", \"street-loyalty-and-danger\", \"violence-and-street-danger\", \"wealth-and-flexing\", \"wealth-and-material-success\"]", "stop_reason": "end_turn", "usage": {"cache_creation": {"ephemeral_1h_input_tokens": 0, "ephemeral_5m_input_tokens": 0}, "cache_creation_input_tokens": 0, "cache_read_input_tokens": 0, "inference_geo": "global", "input_tokens": 1390, "output_tokens": 1201, "output_tokens_details": {"thinking_tokens": 1076}, "service_tier": "batch"}}
+{"id": "group-vulnerable", "text": "[\"disappointment-and-failure\", \"fear-of-losing-love\", \"heartbreak-and-grief\", \"hidden-emotional-struggle\", \"inner-mental-turmoil\", \"loneliness-and-isolation\", \"longing-and-loss\", \"past-trauma-and-healing\", \"self-worth-and-insecurity\", \"vulnerability-and-betrayal\"]", "stop_reason": "end_turn", "usage": {"cache_creation": {"ephemeral_1h_input_tokens": 0, "ephemeral_5m_input_tokens": 0}, "cache_creation_input_tokens": 0, "cache_read_input_tokens": 0, "inference_geo": "global", "input_tokens": 1387, "output_tokens": 677, "output_tokens_details": {"thinking_tokens": 555}, "service_tier": "batch"}}
@@ -0,0 +1,30 @@
+You are selecting thematic keywords that belong to a named
+group.
+
+Input: a JSON object with the name of one group and the
+keywords — the complete set of codes; use these and no
+others:
+
+{
+  "group": "the name of the group",
+  "keywords": ["first-keyword", "second-keyword"]
+}
+
+Task: list every given keyword that belongs to the named
+group, judged by the literal meaning of the keyword itself.
+
+Rules:
+
+- Use only the given keywords, spelled exactly as given.
+- A group may match any number of keywords, including none.
+- Judge each keyword only by the literal meaning of its own
+  wording.
+
+Do not wrap the output in a Markdown code fence.
+The output must be strictly valid JSON.
+
+Output a single JSON array of the keywords that belong to
+the group (an empty array when none belongs), and nothing
+else:
+
+["first-keyword", "second-keyword"]
@@ -0,0 +1,25 @@
+{
+  "script_version": "run_llm.py 3.1.0",
+  "model": "claude-fable-5",
+  "temperature": null,
+  "thinking": null,
+  "max_tokens": 8192,
+  "prompt_path": "../prompts/04-group.md",
+  "prompt_sha256": "50e1d5d6c4115d090ae4ab1e801ea74db0b2dcdd47213ce0feb6749e034a41e9",
+  "input_path": "instance/llm-input-group.jsonl",
+  "input_sha256": "7fd97ae165270cb91461aedf411050892667b928d8584fceffea6f9649e003c7",
+  "item_count": 4,
+  "dry_run": false,
+  "started_at": "2026-08-14T23:36:03+08:00",
+  "batch": {
+    "batch_id": "msgbatch_01C17xW3YBefThTYZ83g7KkL",
+    "submitted_at": "2026-08-14T23:36:04+08:00",
+    "ended_at": "2026-08-14T15:37:24.008527+00:00"
+  },
+  "usage": {
+    "cache_creation_input_tokens": 0,
+    "cache_read_input_tokens": 0,
+    "input_tokens": 5553,
+    "output_tokens": 1891
+  }
+}
@@ -0,0 +1,4 @@
+{"id": "group-women-power", "text": "[\"female-empowerment\", \"women-power\"]", "stop_reason": "end_turn", "usage": {"cache_creation": {"ephemeral_1h_input_tokens": 0, "ephemeral_5m_input_tokens": 0}, "cache_creation_input_tokens": 0, "cache_read_input_tokens": 0, "inference_geo": "global", "input_tokens": 1385, "output_tokens": 62, "output_tokens_details": {"thinking_tokens": 46}, "service_tier": "batch"}}
+{"id": "group-misogyny", "text": "[\"rejection-of-women\"]", "stop_reason": "end_turn", "usage": {"cache_creation": {"ephemeral_1h_input_tokens": 0, "ephemeral_5m_input_tokens": 0}, "cache_creation_input_tokens": 0, "cache_read_input_tokens": 0, "inference_geo": "global", "input_tokens": 1391, "output_tokens": 129, "output_tokens_details": {"thinking_tokens": 117}, "service_tier": "batch"}}
+{"id": "group-masculine", "text": "[\"ambition-and-self-determination\", \"confidence-and-self-assurance\", \"dominance-and-power\", \"family-and-fatherhood\", \"hustle-and-money\", \"relentless-ambition\", \"rivalry-and-superiority\", \"self-confidence-and-braggadocio\", \"showing-off-and-impressing\", \"street-loyalty-and-danger\", \"violence-and-street-danger\", \"wealth-and-flexing\"]", "stop_reason": "end_turn", "usage": {"cache_creation": {"ephemeral_1h_input_tokens": 0, "ephemeral_5m_input_tokens": 0}, "cache_creation_input_tokens": 0, "cache_read_input_tokens": 0, "inference_geo": "global", "input_tokens": 1390, "output_tokens": 684, "output_tokens_details": {"thinking_tokens": 544}, "service_tier": "batch"}}
+{"id": "group-vulnerable", "text": "[\"addiction-and-obsession\", \"alcohol-and-substance-abuse\", \"clinging-to-love\", \"disappointment-and-failure\", \"fear-of-losing-love\", \"heartbreak-and-grief\", \"hidden-emotional-struggle\", \"inner-mental-turmoil\", \"loneliness-and-isolation\", \"longing-and-loss\", \"past-trauma-and-healing\", \"self-worth-and-insecurity\", \"vulnerability-and-betrayal\"]", "stop_reason": "end_turn", "usage": {"cache_creation": {"ephemeral_1h_input_tokens": 0, "ephemeral_5m_input_tokens": 0}, "cache_creation_input_tokens": 0, "cache_read_input_tokens": 0, "inference_geo": "global", "input_tokens": 1387, "output_tokens": 1016, "output_tokens_details": {"thinking_tokens": 859}, "service_tier": "batch"}}
@@ -0,0 +1,30 @@
+You are selecting thematic keywords that belong to a named
+group.
+
+Input: a JSON object with the name of one group and the
+keywords — the complete set of codes; use these and no
+others:
+
+{
+  "group": "the name of the group",
+  "keywords": ["first-keyword", "second-keyword"]
+}
+
+Task: list every given keyword that belongs to the named
+group, judged by the literal meaning of the keyword itself.
+
+Rules:
+
+- Use only the given keywords, spelled exactly as given.
+- A group may match any number of keywords, including none.
+- Judge each keyword only by the literal meaning of its own
+  wording.
+
+Do not wrap the output in a Markdown code fence.
+The output must be strictly valid JSON.
+
+Output a single JSON array of the keywords that belong to
+the group (an empty array when none belongs), and nothing
+else:
+
+["first-keyword", "second-keyword"]
@@ -0,0 +1,25 @@
+{
+  "script_version": "run_llm.py 3.1.0",
+  "model": "claude-fable-5",
+  "temperature": null,
+  "thinking": null,
+  "max_tokens": 8192,
+  "prompt_path": "../prompts/04-group.md",
+  "prompt_sha256": "50e1d5d6c4115d090ae4ab1e801ea74db0b2dcdd47213ce0feb6749e034a41e9",
+  "input_path": "instance/llm-input-group.jsonl",
+  "input_sha256": "7fd97ae165270cb91461aedf411050892667b928d8584fceffea6f9649e003c7",
+  "item_count": 4,
+  "dry_run": false,
+  "started_at": "2026-08-14T23:39:37+08:00",
+  "batch": {
+    "batch_id": "msgbatch_01DveEMYyjCAYe6wxpcCD87V",
+    "submitted_at": "2026-08-14T23:39:43+08:00",
+    "ended_at": "2026-08-14T15:41:46.778015+00:00"
+  },
+  "usage": {
+    "cache_creation_input_tokens": 0,
+    "cache_read_input_tokens": 0,
+    "input_tokens": 5553,
+    "output_tokens": 1975
+  }
+}
@@ -0,0 +1,4 @@
+{"id": "group-women-power", "text": "[\"female-empowerment\", \"women-power\"]", "stop_reason": "end_turn", "usage": {"cache_creation": {"ephemeral_1h_input_tokens": 0, "ephemeral_5m_input_tokens": 0}, "cache_creation_input_tokens": 0, "cache_read_input_tokens": 0, "inference_geo": "global", "input_tokens": 1385, "output_tokens": 77, "output_tokens_details": {"thinking_tokens": 61}, "service_tier": "batch"}}
+{"id": "group-misogyny", "text": "[\"rejection-of-women\"]", "stop_reason": "end_turn", "usage": {"cache_creation": {"ephemeral_1h_input_tokens": 0, "ephemeral_5m_input_tokens": 0}, "cache_creation_input_tokens": 0, "cache_read_input_tokens": 0, "inference_geo": "global", "input_tokens": 1391, "output_tokens": 207, "output_tokens_details": {"thinking_tokens": 195}, "service_tier": "batch"}}
+{"id": "group-masculine", "text": "[\"dominance-and-power\", \"family-and-fatherhood\", \"hustle-and-money\", \"rivalry-and-superiority\", \"self-confidence-and-braggadocio\", \"showing-off-and-impressing\", \"street-loyalty-and-danger\", \"violence-and-street-danger\", \"wealth-and-flexing\"]", "stop_reason": "end_turn", "usage": {"cache_creation": {"ephemeral_1h_input_tokens": 0, "ephemeral_5m_input_tokens": 0}, "cache_creation_input_tokens": 0, "cache_read_input_tokens": 0, "inference_geo": "global", "input_tokens": 1390, "output_tokens": 846, "output_tokens_details": {"thinking_tokens": 740}, "service_tier": "batch"}}
+{"id": "group-vulnerable", "text": "[\"disappointment-and-failure\", \"fear-of-losing-love\", \"heartbreak-and-grief\", \"hidden-emotional-struggle\", \"inner-mental-turmoil\", \"loneliness-and-isolation\", \"longing-and-loss\", \"past-trauma-and-healing\", \"self-worth-and-insecurity\", \"vulnerability-and-betrayal\"]", "stop_reason": "end_turn", "usage": {"cache_creation": {"ephemeral_1h_input_tokens": 0, "ephemeral_5m_input_tokens": 0}, "cache_creation_input_tokens": 0, "cache_read_input_tokens": 0, "inference_geo": "global", "input_tokens": 1387, "output_tokens": 845, "output_tokens_details": {"thinking_tokens": 723}, "service_tier": "batch"}}
@@ -0,0 +1,30 @@
+You are selecting thematic keywords that belong to a named
+group.
+
+Input: a JSON object with the name of one group and the
+keywords — the complete set of codes; use these and no
+others:
+
+{
+  "group": "the name of the group",
+  "keywords": ["first-keyword", "second-keyword"]
+}
+
+Task: list every given keyword that belongs to the named
+group, judged by the literal meaning of the keyword itself.
+
+Rules:
+
+- Use only the given keywords, spelled exactly as given.
+- A group may match any number of keywords, including none.
+- Judge each keyword only by the literal meaning of its own
+  wording.
+
+Do not wrap the output in a Markdown code fence.
+The output must be strictly valid JSON.
+
+Output a single JSON array of the keywords that belong to
+the group (an empty array when none belongs), and nothing
+else:
+
+["first-keyword", "second-keyword"]