可重複的查核流程A repeatable verification workflow

原則Principles

  • 查核不是「再問一次 AI」，是拿 AI 看不到的東西——你的 jamovi 介面、你的原始資料、你的研究設計——去對。
  • 兩條路線難度不對等：Module Guider 的輸出有有限集合可核（選單路徑）； R code tutor 的輸出只能靠流程（讀碼、跑碼、跟 GUI 對照）。
  • 每次查核產出一筆「協作日誌」（見第六課），這是可教、可重複、可交作業的單位。
  • Verification is not “asking the AI again.” It means checking what the AI said against things it cannot see: your jamovi interface, your raw data, your research design.
  • The two paths are not equally hard to verify: the Module Guider’s output can be checked against a finite, verifiable set (menu paths); the R code tutor’s output can only be checked through a process (read the code, run it, cross-check against the GUI).
  • Every verification pass produces one “collaboration log” entry (see Lesson 6). You can teach that unit, repeat it, and turn it into an assignment.

流程步驟Workflow steps

送出前   P0 寫判準（我會接受什麼答案）           ← 第 3 課
         P1 記下摘要裡的關鍵數字（n、mean、遺漏）

收到後   V1 路徑核對（Guider）：每條 Analyses ▸ … 在選單點得到？逐條打勾
         V2 數字核對：回覆引用的數字 = P1 記下的？
         V3 範圍核對：回覆有沒有宣稱它「看不到」的事（常態性、離群值、原始列）？有 → 標「越界」
         V4 程式碼核對（R tutor）：
              4a 讀：只用 data？library() 都在 rj_environment？無 setwd／路徑／system？
              4b 跑：貼進 Rj，記錄輸出或錯誤原文
              4c 對：同一分析在 jamovi GUI 跑一次，關鍵數字一致？（Syntax Mode 的 jmv:: 呼叫可作第二對照）
         V5 交叉：換 persona 或換模型再問一次，結論一致？路徑仍須各自核對

結案     C1 對照 P0 判準：solved / reopen
         C2 填協作日誌（問題｜判準｜回覆摘要｜查核結果）
Before sending   P0 Write your criteria (what answer you will accept)   <- Lesson 3
                 P1 Note the key numbers in the summary (n, mean, missing)

After receiving  V1 Path check (Guider): can every Analyses ▸ … path be clicked in the menu?
                    Tick each one
                 V2 Number check: do the numbers the reply cites match what you noted in P1?
                 V3 Scope check: does the reply claim things it cannot see (normality, outliers, raw rows)?
                    If so, flag it as overreaching
                 V4 Code check (R tutor):
                      4a Read: uses only data? every library() is in rj_environment? no setwd, file paths, or system?
                      4b Run: paste it into Rj and record the output or the exact error
                      4c Cross-check: run the same analysis in the jamovi GUI; do the key numbers match?
                         (the jmv:: call from Syntax Mode can serve as a second reference)
                 V5 Cross-check: ask again with another persona or model; do the conclusions agree?
                    Check each reply's paths separately

Closing          C1 Compare with the P0 criteria: solved / reopen
                 C2 Fill in the collaboration log (question | criteria | reply summary | verification result)

兩份查核清單Two checklists