R code tutor 查核清單R Code Tutor Verification Checklist

本站有兩份查核清單，這是其中第二份：涵蓋 verify/workflow.qmd 流程裡的 P0–P1、V2–V4、V5、C1–C2 步驟，設計成可列印的一頁。適用對象是使用 R code tutor（askllmr）取得 R 程式碼的情境；若你用的是 Module Guider（askllm），請改用 checklist-guider.qmd。

This is the second of the two verification checklists on this site. It covers steps P0–P1, V2–V4, V5, and C1–C2 of the workflow in verify/workflow.qmd, laid out as a single printable page. Use it when you asked the R code tutor (askllmr) for R code. If you used the Module Guider (askllm) instead, use checklist-guider.qmd.

與 checklist-guider.qmd 最大的差異在於：R code tutor 的答案空間不是有限 集合（不像選單路徑可以逐條核對），R 是 Turing-complete，沒有任何掃描能證明 一段任意生成的程式碼「如它所說地運作」。所以這裡沒有 V1 路徑核對，取而代之 的是 V4 程式碼核對，拆成讀（4a）、跑（4b）、對（4c）三段——這是流程性 的查核，不是機械性的查核。詳見 askLLM/docs/LIMITATIONS.zh-TW.md〈兩個分析， 兩種查證方式〉一節。

The key difference from checklist-guider.qmd is that the R code tutor’s output space is not a finite, checkable set the way menu paths are. R is Turing-complete, so no scan can prove that an arbitrary piece of generated code actually “works as claimed.” So there is no V1 path check here. Its place is taken by V4, the code check, split into read (4a), run (4b), and cross-check (4c). This is a procedural check, not a mechanical one. See the “Two analyses, two ways to verify” section of askLLM/docs/LIMITATIONS.zh-TW.md for background.

送出前Before you submit

• P0 我已經寫下「我會接受什麼答案」（例如：程式碼必須只用 data 這個物件名稱、必須產出我要的統計量、不得需要我自己補檔案路徑）。
• P1 我已經記下摘要裡的關鍵數字（每一組的 n、mean、sd、遺漏數）， 這些數字之後要拿來跟程式碼的輸出核對。

• P0 I have written down what answer I will accept (e.g. “the code must use only the object name data, must produce the statistic I asked for, and must not require me to fill in a file path myself”).
• P1 I have written down the key numbers from the summary (n, mean, sd, missing count per group) so I can check the code’s output against them later.

收到回覆後After you receive a reply

• V2 數字核對 — 執行後的輸出數字（n、mean、sd 等），跟我在 P1 記下的、以及在 Descriptives 面板看到的一致。
• V3 範圍核對 — 回覆有沒有宣稱它「看不到」的東西（例如常態性、離 群值、原始資料列的內容）？如果有，這是「越界」，代表回覆超出摘要統計 能支持的範圍，應打折扣看待。
• V4a 讀 — 逐行讀完整段程式碼，確認： - 資料只用 data 這個物件名稱，沒有另外指定其他資料來源； - 每一個 library(...) 引用的套件，都出現在 askLLM 附給模型的 <rj_environment> 套件清單內，或屬於 base R； - 沒有出現「紅旗指令」（見下方紅旗字串表）。 只要有一項沒通過，就先不要執行，回頭修改問題或換一個 persona 再問一次。
• V4b 跑 — 把程式碼原封不動貼進 Rj Editor（Analyses ▸ R ▸ Rj ▸ Rj Editor），按下 Run，把輸出或錯誤訊息原文記下來（成功的輸出 與失敗的錯誤訊息都要記，不要只記成功的）。
• V4c 對 — 用另一個獨立管道重跑同一個分析，比較關鍵數字是否一致： 改用 jamovi GUI 對應的分析模組跑一次；或在該分析的 Syntax Mode 中查看 jamovi 自己產生的 jmv:: 呼叫，作為第二層獨立對照。
• V5 交叉核對 — 換一個 persona（Consultant / Tutor / Explainer） 或換一個模型，針對同一個 What should the R code do? 問題再問一次； 兩次程式碼的邏輯與輸出是否一致？（V4a／V4b 仍須對第二份程式碼各自 重跑一次，不能因為兩次都「看起來合理」就省略讀碼與執行。）

• V2 Number check — The numbers produced by running the code (n, mean, sd, etc.) match what I wrote down in P1 and what I see in the Descriptives panel.
• V3 Scope check — Does the reply make claims about things it cannot actually see (e.g. normality, outliers, the content of raw data rows)? If so, flag this as “overreaching.” The claim exceeds what summary statistics can support, and you should discount it.
• V4a Read — Read the entire code block line by line and confirm: - the data is referenced only through the object name data, with no other data source specified; - every library(...) call names a package that appears in the <rj_environment> package list askLLM sent the model, or is base R; - none of the red-flag instructions below appear. If any of these fails, do not run the code. Rephrase the question or try a different persona instead.
• V4b Run — Paste the code verbatim, unmodified into the Rj Editor (Analyses ▸ R ▸ Rj ▸ Rj Editor), press Run, and record the exact output or error message. Record both a successful output and a failed error message, not only the successes.
• V4c Cross-check — Re-run the same analysis through an independent channel and compare the key numbers: run the equivalent analysis through the jamovi GUI; or check that analysis’s Syntax Mode for the jmv:: call jamovi generates itself, as a second, independent comparison.
• V5 Cross-check (persona/model) — Ask the same “What should the R code do?” question again with a different persona (Consultant / Tutor / Explainer) or a different model. Do the two pieces of code agree in logic and output? (Still repeat V4a/V4b on the second piece of code — two replies “looking reasonable” is not a substitute for reading and running each one.)

結案Closing out

• C1 對照 P0 寫下的判準：solved（全部命中，程式碼可用）還是 reopen （任一項沒通過，需要換個問法或換個 persona 再試）。
• C2 填寫協作日誌四欄：問題 / 判準 / 回覆摘要 / 查核結果。

• C1 Compare the reply against the P0 criteria, and reach one of two outcomes. Solved means everything checks out and the code is usable. Reopen means something failed, so rephrase the question or try a different persona.
• C2 Fill in the four-column collaboration log: question / criteria / reply summary / verification result.

紅旗字串表（4a 逐行核對用）Red-flag string table (for V4a)

askLLM 的 system prompt 明文禁止模型在 R code tutor 的回覆裡使用下列指令 （見 askLLM/R/llm-adapter.R 第 172–177 行）。若程式碼裡出現任何一項，代表 回覆違反了 askLLM 自己設定的邊界，應標記為「不可執行」並回頭修改問題：

askLLM’s system prompt explicitly forbids the model from using the following in an R code tutor reply (see askLLM/R/llm-adapter.R, lines 172–177). If any of these appear in the code, the reply has violated askLLM’s own stated boundary. Flag it as “do not run” and go back and rephrase the question:

紅旗Red flag	為什麼Why
install.packages(...)	模型不能替你安裝套件；Rj 只用已裝好的套件。The model cannot install packages for you; Rj only has what is already installed.
read.csv(...)／任何檔案讀取函式/ any file-reading function	不得建議讀取檔案；資料一律用 data 這個既有物件。No file reads are allowed; the data must always come from the existing data object.
setwd(...)	不得變更工作目錄。The working directory must not be changed.
任何絕對路徑（如 C:/...、/Users/...）Any absolute file path (e.g. C:/..., /Users/...)	路徑因人而異，不可能通用，且暗示前一條「檔案讀取」紅旗。Paths are machine-specific and can’t generalize; this usually co-occurs with the file-read red flag above.
system(...)	不得呼叫系統層級指令。System-level shell calls are not allowed.
library(套件)，但該套件不在 <rj_environment> 清單內、也非 base R	模型只能引用你 Rj 裡實際裝了的套件；引用未裝套件的程式碼一執行就會出現 there is no package called '...'。library(package) where the package is not in the <rj_environment> list and is not base R — the model may only cite packages actually installed in your Rj; code citing an uninstalled package will fail immediately with there is no package called '...'.

提醒：<rj_environment> 這個套件清單只有在你的 jamovi 裝有 Rj、且 askLLM 掃描到套件時才會附上。若你的 jamovi 沒裝 Rj，askLLM 會改教你 用 Syntax Mode，而不是硬生出一段你根本跑不了的程式碼；這種情況下本清單 的 4a／4b 步驟不適用，因為根本沒有 Rj Editor 可以貼上執行。

Note: the <rj_environment> package list is only attached when your jamovi has Rj installed and askLLM was able to scan its packages. Without Rj installed, askLLM should instead teach you to use Syntax Mode rather than generating code you cannot actually run. Steps 4a/4b of this checklist then do not apply. There is no Rj Editor to paste into.

相關頁面Related pages

• 完整流程說明：verify/workflow.qmd
• Module Guider 版查核清單：checklist-guider.qmd
• 已知錯誤範例與「兩個分析，兩種查證方式」說明： askLLM/docs/LIMITATIONS.zh-TW.md（母專案文件，不在本站範圍內修改）
• 找碴練習（含程式碼紅旗題）：spot-the-error.qmd

• Full workflow: verify/workflow.qmd
• Module Guider checklist: checklist-guider.qmd
• Recorded real errors and the “two analyses, two ways to verify” discussion: askLLM/docs/LIMITATIONS.zh-TW.md (parent-project document, not modified within this site)
• Spot-the-error practice (includes code red-flag questions): spot-the-error.qmd