24 KiB
Grandma Clinic — AI Bugs Made Simple(問題地圖 1–16)
為什麼會有這個頁面
大多數人都是在模型已經講完話之後才修 AI bug。接著加上 patch、reranker、或 regex。結果同一種失敗之後會換個樣子再回來。
WFGY 在「輸出之前」就裝上一道語義防火牆。
它會先檢查語義場。如果狀態不穩,就進入循環、收窄、或重置。只有穩定的狀態才被允許發言。只要把失敗模式映射完成,它就會一直保持被修復的狀態。
30 秒上手用法
- 滾動到最像你案例的編號。
- 讀「奶奶故事」。如果對得上,複製下面的醫生提示詞。
- 把提示詞貼到 Dr. WFGY 與醫生對話。
連結: Dr. WFGY in ChatGPT Room - 你會同時拿到「簡單修法」與「專業修法」。不需要 SDK。
不確定從哪開始? 先用 Beginner Guide 快速定位你的問題,跑完第一個安全修復,再進診所。
快速連結
如果你的整個 stack 連開都開不起來,先看這三個:
No.14 Bootstrap Ordering
No.15 Deployment Deadlock
No.16 Pre-deploy Collapse
每段的格式規則
• 一般文字 = 奶奶故事、比喻對應、**奶奶防呆(輸出前)**含映射、Minimal fix 與提示詞。
• Pro Zone = 可展開區塊:準確症狀、技術關鍵與參考連結。
No.1 Hallucination & Chunk Drift — 奶奶:拿錯食譜
Grandma story
You ask for the cabbage recipe. I hand you a random page from a different cookbook because its picture looks similar.
Metaphor mapping
- 漂亮圖片 = token 表面匹配
- 錯的食譜書 = 錯誤來源
- 好聽口氣 = 沒證據的自信語氣
Grandma fix(輸出前)— 映射
- 先把食譜卡 擺上桌 = citation-first policy
- 標出使用的書與頁碼 = 檢索追蹤(ID/頁碼)
- 下鍋前先核對「cabbage」= 查詢–來源語義檢查(ΔS gate)
Minimal fix(grandma)
Do not taste anything until the recipe card is on the table.
Doctor prompt:
please explain No.1 Hallucination & Chunk Drift in grandma mode, then show me the minimal WFGY fix and the exact reference link
Pro Zone
Real scene
Bad OCR 或不良分塊造成碎片。檢索挑到高 cosine 但語義錯誤的鄰居。模型說得很順卻沒有引用。
Technical keys
- 開啟 citation-first policy
- 加上檢索追蹤:ID 與來源頁
- 檢查分塊規則與表格處理
- 確認來源後再加最小 reranker
Reference:
Hallucination & Chunk Drift → https://github.com/onestardao/WFGY/blob/main/ProblemMap/hallucination.md
No.2 Interpretation Collapse — 奶奶:把糖當鹽
Grandma story
You found the right page but misread the steps. Sugar replaced with salt. The dish fails even with the correct book open.
Metaphor mapping
- 正確頁面 = 正確 chunk
- 讀錯步驟 = 推理崩壞
- 吃起來不對 = 有檢索仍答錯
Grandma fix(輸出前)— 映射
- 每步 慢讀並唸出來 = λ_observe 中途檢查點
- 倒料前先劃線標示數量 = 符號/約束锚定
- 味道跑掉就 暫停重讀 = BBCR 受控重置
Minimal fix(grandma)
Read slowly. When unsure, stop and ask a checkpoint.
Doctor prompt:
please explain No.2 Interpretation Collapse in grandma mode, then apply a minimal WFGY checkpoint plan
Pro Zone
Real scene
檢索後答案漂移。模型在正確上下文中推理,卻在鏈中途失去結構。
Technical keys
- 量測 ΔS(提示 vs 答案)
- 插入 λ_observe 檢查點
- 若仍漂移,做 BBCR 控制重置
- 完成前 Coverage ≥ 0.70
Reference:
Interpretation Collapse → https://github.com/onestardao/WFGY/blob/main/ProblemMap/retrieval-collapse.md
No.3 Long Reasoning Chains — 奶奶:越逛越忘
Grandma story
You go to market A, then B, then C, and forget why you left home.
Metaphor mapping
- 好多站 = 推理步驟太長
- 忘了目標 = 情境漂移
- 買對物品、做錯菜 = 跟目標不符
Grandma fix(輸出前)— 映射
- 購物清單把 主菜寫最上面 = 目標锚(goal anchor)
- 每兩條街對一次清單 = 循環+檢查點
- 袋中物 vs 清單比對 = Coverage 門檻
Minimal fix(grandma)
Write the shopping list and check it every two streets.
Doctor prompt:
please explain No.3 Long Reasoning Chains in grandma mode and show the smallest loop + checkpoint pattern
Pro Zone
Real scene
多步計畫走偏。早期決策沒有回檢。最後答案看似完整卻偏離目標。
Technical keys
- 明確定義目標锚
- 用 λ_diverse 比較 3+ 路徑
- 限制 CoT 變異並修剪離題分支
- 每輪對目標锚重評分
Reference:
Long Reasoning Chains → https://github.com/onestardao/WFGY/blob/main/ProblemMap/context-drift.md
No.4 Bluffing / Overconfidence — 奶奶:沒卡別端菜
Grandma story
A charming waiter serves a dish without showing the recipe card. Sounds right, tastes wrong.
Metaphor mapping
- 自信語氣 = 流利自然語言
- 沒食譜卡 = 無證據
- 禮貌微笑 = 道歉不修復
Grandma fix(輸出前)— 映射
- 「先看卡」= 證據先於答案
- 沒卡退回去 = 拒絕無根答案
- 記錄「哪張卡做哪道菜」= 可追蹤日誌
Minimal fix(grandma)
Ask for the card first. If none, send the dish back.
Doctor prompt:
please explain No.4 Bluffing in grandma mode, then enforce 'card first' with a minimal WFGY guardrail
Pro Zone
Real scene
自然語言聽起來很對但其實錯。缺乏可追溯路徑。模型拒絕驗證。
Technical keys
- Citation-first policy
- 拒絕無根斷言
- 確認來源後再做最小 reranker
- 紀錄 Coverage 與 ΔS
Reference:
Bluffing / Overconfidence → https://github.com/onestardao/WFGY/blob/main/ProblemMap/bluffing.md
No.5 Semantic ≠ Embedding — 奶奶:胡椒名同味不同
Grandma story
White pepper and black pepper. Same word “pepper,” completely different flavor.
Metaphor mapping
- 同詞不同義 = 表面 token 重疊
- 風味不同 = 語義不相等
- 分數高仍錯 = 高相似≠同意思
Grandma fix(輸出前)— 映射
- 兩個都聞/嚐 = 度量健檢(metric sanity)
- 不混標籤不清的罐子 = 空間正規化+大小寫一致
- 留一口 標準對照湯 = 小型真值樣例
Minimal fix(grandma)
Taste both peppers before cooking.
Doctor prompt:
please explain No.5 Semantic ≠ Embedding in grandma mode and give me the minimal metric audit plan
Pro Zone
Real scene
未正規化向量、混用模型向量、大小寫與分詞不一致,導致選到語義不等價鄰居。
Technical keys
- 向量正規化
- 驗證度量空間與維度
- 對齊分詞與大小寫
- 先通過度量稽核再談混合檢索
Reference:
Semantic ≠ Embedding → https://github.com/onestardao/WFGY/blob/main/ProblemMap/embedding-vs-semantic.md
No.6 Logic Collapse & Recovery — 奶奶:死巷一直撞
Grandma story
You keep taking the same dead-end alley. Step back, pick a new street, and try again.
Metaphor mapping
- 死胡同 = 無效迴圈
- 後退 = 受控重置
- 換路 = 替代路徑
Grandma fix(輸出前)— 映射
- 撞牆兩次就 回頭 = ΔS 連續高就 BBCR 重置
- 換 下一條街 試 = 替代候選路徑
- 手上拿地圖 = 狀態锚+目標提醒
Minimal fix(grandma)
If lost twice, stop and change route.
Doctor prompt:
please explain No.6 Logic Collapse in grandma mode, then show BBCR reset + λ\_observe checkpoints
Pro Zone
Real scene
推理卡死在環或淺分支。缺乏偵測與恢復機制。
Technical keys
- 每步量測 ΔS
- λ_observe 鏈中落地
- ΔS 居高不下則 BBCR
- 只接受收斂 λ 與 Coverage ≥ 0.70
Reference:
Logic Collapse & Recovery → https://github.com/onestardao/WFGY/blob/main/ProblemMap/logic-collapse.md
No.7 Memory Breaks Across Sessions — 奶奶:記在錯抽屜
Grandma story
You promise to remember the family recipe, then next week you act like we never talked.
Metaphor mapping
- 忘了鍋上的刮痕 = 狀態遺失
- 每次都是新廚房 = 無連續性
- 一問再問 = 用戶疲勞
Grandma fix(輸出前)— 映射
- 寫在 標籤卡 上 = 穩定記憶結構/state keys
- 永遠放 同一個抽屜 = 寫讀順序防護
- 卡上貼小照片 = 低 ΔS 範例庫
Minimal fix(grandma)
Write notes on a card and keep it in the same drawer.
Doctor prompt:
please explain No.7 Memory Breaks in grandma mode and show the smallest stable memory routine
Pro Zone
Real scene
Session 狀態、锚點、合約未持久或無追蹤,導致無聲上下文遺失。
Technical keys
- 穩定記憶綱要與 state keys
- 寫讀順序防護
- 小型示例庫處理低 ΔS 案例
- 以 ID 可追蹤的檢索
Reference:
Memory Coherence → https://github.com/onestardao/WFGY/blob/main/ProblemMap/memory-coherence.md
No.8 Debugging is a Black Box — 奶奶:空白卡片
Grandma story
You tell me “trust me, it works.” I ask “show me which page you used.” You shrug.
Metaphor mapping
- 盲煮 = 無追蹤
- 「我記得」= 無法驗證
- 不能重做 = 不可重現
Grandma fix(輸出前)— 映射
- 食譜卡 釘在爐子旁 = 答案同時呈現來源
- 標上 頁碼 = ID/行號追蹤
- 留一張「我怎麼煮的」小紙條 = 最小可重現管線
Minimal fix(grandma)
Pin the recipe card next to the stove.
Doctor prompt:
please explain No.8 Debugging Black Box in grandma mode and add a tiny traceability schema
Pro Zone
Real scene
沒有 ID 或來源行,難以證明哪個 chunk 產生答案,修復全靠猜。
Technical keys
- 檢索可追蹤(IDs)
- 紀錄 query、chunk IDs、接受度指標
- 最小可重現管線
- 最終答案前先檢查「是否有來源」
Reference:
Retrieval Traceability → https://github.com/onestardao/WFGY/blob/main/ProblemMap/retrieval-traceability.md
No.9 Entropy Collapse — 奶奶:一鍋灰色大雜燴
Grandma story
Too many voices in one room. Everyone talks. Nobody listens. The dish becomes mush.
Metaphor mapping
- 噪音 = 熵過載
- 融化的注意力 = 無結構
- 一鍋灰泥 = 內在不一致
Grandma fix(輸出前)— 映射
- 關小火、一步一步煮 = 降低步寬
- 先分好 角色/關係/限制 碗 = 锚定實體與約束
- 上桌前要先嚐 = 接受門檻(ΔS、Coverage)
Minimal fix(grandma)
Lower the heat and separate steps.
Doctor prompt:
please explain No.9 Entropy Collapse in grandma mode and show a minimal stability recipe
Pro Zone
Real scene
注意力擴散,路徑混雜。表面流暢但內部矛盾。
Technical keys
- 降低步寬
- 锚定實體、關係、約束
- 夾制變異並要求 Coverage
- 最終輸出前設接受目標
Reference:
Entropy Collapse → https://github.com/onestardao/WFGY/blob/main/ProblemMap/entropy-collapse.md
No.10 Creative Freeze — 奶奶:湯可吃但好無聊
Grandma story
You only follow the recipe word by word. The soup is edible, never memorable.
Metaphor mapping
- 沒加香料 = 字面輸出
- 不試味 = 低探索
- 平淡無奇 = 無趣答案
Grandma fix(輸出前)— 映射
- 並排試 兩三種安全調味 = λ_diverse 候選
- 全部對著同一張成品照比較 = 共享锚評分
- 味道在「微~中等」區間 = 受控熵窗口
Minimal fix(grandma)
Taste and adjust within a safe range.
Doctor prompt:
please explain No.10 Creative Freeze in grandma mode and give the smallest safe-exploration pattern
Pro Zone
Real scene
模型逃避多樣候選,全部收斂成平庸答案。
Technical keys
- λ_diverse 產生答案集合
- 受控熵窗口
- 以同一锚比較候選
- ΔS 保持在可接受範圍
Reference:
Creative Freeze → https://github.com/onestardao/WFGY/blob/main/ProblemMap/creative-freeze.md
No.11 Symbolic Collapse — 奶奶:看字會算數不行
Grandma story
You can read the storybook but panic when you see fractions and tables.
Metaphor mapping
- 文字 OK = 自然語言沒問題
- 符號可怕 = 數學或表格失靈
- 故事好聽、數學錯 = 結構被壓平成散文
Grandma fix(輸出前)— 映射
- 把 數字放在框裡 = 獨立符號通道
- 表格別改寫成散文 = 保留區塊
- 喊出單位(grams, tsp)= 運算子/單位锚定
- 先做一小口試煉 = 微型證明/例子
Minimal fix(grandma)
Keep the story but show the table step by step.
Doctor prompt:
please explain No.11 Symbolic Collapse in grandma mode and show me a minimal symbol-first routine
Pro Zone
Real scene
公式、運算子、程式碼區塊、標題被壓平成散文。答案看似順卻錯。
Technical keys
- 獨立符號通道
- 保留 code/table 區塊
- 锚定運算子與單位
- 以小證明或例子驗證
Reference:
Symbolic Collapse → https://github.com/onestardao/WFGY/blob/main/ProblemMap/symbolic-collapse.md
No.12 Philosophical Recursion — 奶奶:無限為什麼
Grandma story
Asking “why” about “why” about “why.” You spin in circles and never cook.
Metaphor mapping
- 無盡鏡像 = 自我指涉
- 螺旋碗 = 悖論陷阱
- 冷灶台 = 沒有最終答案
Grandma fix(輸出前)— 映射
- 寫下 頂層問題 便利貼 = 外框/锚
- 只允許 N 次 why(如 2) = 遞迴停止規則
- 收尾一定要有 實例/引用 = 落地要求
Minimal fix(grandma)
Set a top question and limit how many mirrors you look into.
Doctor prompt:
please explain No.12 Philosophical Recursion in grandma mode and give me a minimal boundary plan
Pro Zone
Real scene
自指與悖論問題使推理無限打轉。
Technical keys
- 定義锚與外框
- ε_resonance 作領域和諧
- 遞迴停止條件
- 需要有例子或引用支撐
Reference:
Philosophical Recursion → https://github.com/onestardao/WFGY/blob/main/ProblemMap/philosophical-recursion.md
No.13 Multi-Agent Chaos — 奶奶:廚房拔河
Grandma story
Two cooks share one kitchen. One adds salt while the other removes it. The soup never stabilizes.
Metaphor mapping
- 共用廚房 = 共用記憶
- 交叉便條 = 角色飄移
- 鹽的拉扯 = 記憶覆寫
Grandma fix(輸出前)— 映射
- 每位廚師各有 署名卡 = 角色與 state keys
- 便條分 不同抽屜 = 所有權與欄柵
- 爐台使用有 計時 = 工具超時/選擇閘
Minimal fix(grandma)
Give each cook a clear card and a separate drawer.
Doctor prompt:
please explain No.13 Multi-Agent Chaos in grandma mode and set a tiny role + memory fence plan
Pro Zone
Real scene
多 agent 互相覆寫狀態或混淆角色。沒有單一真相來源。
Technical keys
- 角色/記憶欄柵
- State keys 與所有權
- 工具超時與選擇閘
- 跨 agent 追蹤
Reference:
Multi-Agent Problems → https://github.com/onestardao/WFGY/blob/main/ProblemMap/Multi-Agent_Problems.md
No.14 Bootstrap Ordering — 奶奶:冷鍋打蛋
Grandma story
You try to fry eggs before turning on the stove. Of course nothing happens.
Metaphor mapping
- 冷鍋 = 服務未就緒
- 先打蛋 = 依賴尚未啟動就呼叫
- 時序燒焦 = 少了熱身步驟
Grandma fix(輸出前)— 映射
- 先開火 → 鍋熱 → 再打蛋 = readiness probes/啟動順序
- 先把油與鍋預熱 = 快取/索引暖機
- 檢查瓦斯與火柴 = 密鑰/權限檢查
Minimal fix(grandma)
Start the fire, heat the pan, then crack the eggs.
Doctor prompt:
please explain No.14 Bootstrap Ordering in grandma mode and give me the smallest boot checklist
Pro Zone
Real scene
服務在相依尚未就緒時啟動。首呼失敗、快取冰冷、密鑰缺失。
Technical keys
- 啟動順序與就緒探針
- 快取暖機與索引切換
- 密鑰檢查與健康閘
- 上公有流量前先走影子流量
Reference:
Bootstrap Ordering → https://github.com/onestardao/WFGY/blob/main/ProblemMap/bootstrap-ordering.md
No.15 Deployment Deadlock — 奶奶:你先我先卡門口
Grandma story
Two people at a narrow doorway say “you first.” “No, you first.” They block the door together.
Metaphor mapping
- 窄門 = 共用資源
- 互相禮讓 = 互鎖等待
- 門口堵塞 = 系統凍結
Grandma fix(輸出前)— 映射
- 指定先後順序 = total order/priority
- 側門繞過 = fallback path
- 禮貌倒數 = timeouts/backoff
Minimal fix(grandma)
Decide who goes first, or open a side door.
Doctor prompt:
please explain No.15 Deployment Deadlock in grandma mode and show the smallest unlock plan
Pro Zone
Real scene
migrator 等 writer;writer 等 migrator;沒有超時,整體停滯。
Technical keys
- 打破相依循環
- 超時與退避
- 臨時唯讀模式
- 發佈閘與回歸檢查
Reference:
Deployment Deadlock → https://github.com/onestardao/WFGY/blob/main/ProblemMap/deployment-deadlock.md
No.16 Pre-deploy Collapse — 奶奶:第一鍋就糊了
Grandma story
First pot burns because you forgot to wash it and check the gas.
Metaphor mapping
- 髒鍋 = 舊版本/索引偏移
- 沒檢查瓦斯 = 秘密或權限缺失
- 第一口就焦 = 首次呼叫崩潰
Grandma fix(輸出前)— 映射
- 先洗鍋與工具 = 版本釘住/乾淨狀態
- 試火 = 環境與 secrets 的 preflight
- 先煎 一顆小蛋 = 小流量金絲雀
Minimal fix(grandma)
Wash the pot, test the flame, cook a tiny egg before guests arrive.
Doctor prompt:
please explain No.16 Pre-deploy Collapse in grandma mode and give me the smallest preflight checklist
Pro Zone
Real scene
版本偏移、環境變數或 secrets 缺失、向量索引首批為空、分析器錯誤,導致第一個線上請求崩潰。
Technical keys
- Preflight 合約檢查
- 版本釘住與模型鎖定
- 向量索引建好再切換
- 金絲雀在最小流量上
Reference:
Pre-deploy Collapse → https://github.com/onestardao/WFGY/blob/main/ProblemMap/predeploy-collapse.md
修好一個之後會怎樣
不是無止境貼 OK 繃。你要設定並維持 接受標準:
- ΔS ≤ 0.45
- Coverage ≥ 0.70
- λ 狀態收斂
- 最終輸出前必須有來源
當新 bug 出現,把它映射到編號,套一次修法,它就會一直被修好。這就是語義防火牆的目的。
一句話醫生提示詞
如果不確定是哪一號:
i’ve uploaded TXT OS / WFGY notes.
which Problem Map number matches my issue?
explain using grandma mode, then give the minimal fix and the reference page.
🔗 一分鐘快速下載
| 工具 | 連結 | 三步驟設定 |
|---|---|---|
| WFGY 1.0 PDF | Engine Paper | 1️⃣ 下載 · 2️⃣ 上傳到你的 LLM · 3️⃣ 詢問 “Answer using WFGY + <your question>” |
| TXT OS(純文字作業系統) | TXTOS.txt | 1️⃣ 下載 · 2️⃣ 貼到任一 LLM 對話 · 3️⃣ 打字 “hello world” — OS 立刻開機 |
🧭 繼續探索
| 模組 | 說明 | 連結 |
|---|---|---|
| WFGY Core | WFGY 2.0 引擎上線:完整符號推理架構與數學堆疊 | View → |
| Problem Map 1.0 | 起始 16 模式診斷與符號修復框架 | View → |
| Problem Map 2.0 | 以 RAG 為中心的失敗樹、模組化修復與管線 | View → |
| Semantic Clinic Index | 擴充故障目錄:提示注入、記憶錯誤、邏輯漂移 | View → |
| Semantic Blueprint | 以層為基礎的符號推理與語義調變 | View → |
| Benchmark vs GPT-5 | 用完整 WFGY 推理套件壓測 GPT-5 | View → |
| 🧙♂️ Starter Village 🏡 | 新手入口,巫師帶你逛符號世界 | Start → |
👑 早期 Stargazers:名人堂 —
從第一天就支持 WFGY 的工程師、駭客與開源夥伴。
⭐ WFGY Engine 2.0 已解鎖。⭐ Star 這個 repo 幫更多人找到它,並解鎖 Unlock Board。
















