mirror of
https://github.com/onestardao/WFGY.git
synced 2026-05-02 13:31:07 +00:00
105 lines
3.6 KiB
Markdown
105 lines
3.6 KiB
Markdown
# 📒 Problem · Bluffing — The Model Pretends to Know
|
||
|
||
Large language models often answer **even when no supporting knowledge exists**.
|
||
This “confident nonsense” is lethal in support bots, policy tools, or any high‑stakes domain.
|
||
WFGY kills bluffing by treating “I don’t know” as a valid, traceable state.
|
||
|
||
---
|
||
|
||
## 🤔 Why Do Models Bluff?
|
||
|
||
| Root Cause | Practical Outcome |
|
||
|------------|------------------|
|
||
| **No Uncertainty Gauge** | LLMs lack an internal “stop” threshold |
|
||
| **Fluency ≠ Truth** | High token probability sounds plausible, not factual |
|
||
| **No Self‑Validation** | Model can’t verify its logic path |
|
||
| **RAG Adds Content, Not Honesty** | Retriever fills context but can’t force humility |
|
||
|
||
---
|
||
|
||
## 🛡️ WFGY Anti‑Bluff Stack
|
||
|
||
| Mechanism | Action |
|
||
|-----------|--------|
|
||
| **ΔS Stress + λ_observe** | Detects chaotic or divergent logic flow |
|
||
| **BBCR Collapse–Rebirth** | Halts output, re‑anchors to last valid Tree node |
|
||
| **Allowed “No‑Answer”** | Model may ask for more context or admit unknowns |
|
||
| **User‑Aware Fallback** | Suggests doc upload or clarification instead of guessing |
|
||
|
||
```text
|
||
"This request exceeds current context.
|
||
No references found. Please add a source or clarify intent."
|
||
````
|
||
|
||
---
|
||
|
||
## ✍️ Quick Test (90 sec)
|
||
|
||
```txt
|
||
1️⃣ Start
|
||
> Start
|
||
|
||
2️⃣ Ask an edge‑case question
|
||
> "Is warranty coverage for lunar colonies mentioned anywhere?"
|
||
|
||
Watch WFGY:
|
||
• ΔS spikes → λ_observe chaotic
|
||
• BBCR halts bluffing
|
||
• Returns a clarification prompt
|
||
```
|
||
|
||
---
|
||
|
||
## 🔬 Sample Output
|
||
|
||
```txt
|
||
No mapped content on lunar‑colony warranties.
|
||
Add a relevant policy document or refine the question.
|
||
```
|
||
|
||
Zero bluff. Full epistemic honesty.
|
||
|
||
---
|
||
|
||
## 🛠 Module Cheat‑Sheet
|
||
|
||
| Module | Role |
|
||
| ----------------- | ------------------------------------- |
|
||
| **ΔS Metric** | Early bluff warning |
|
||
| **λ\_observe** | Flags chaos states |
|
||
| **BBCR** | Stops & resets logic |
|
||
| **Semantic Tree** | Stores last valid anchor |
|
||
| **BBAM** | Lowers overconfident attention spikes |
|
||
|
||
---
|
||
|
||
## 📊 Implementation Status
|
||
|
||
| Feature | State |
|
||
| --------------------------- | -------- |
|
||
| Bluff detection | ✅ Stable |
|
||
| BBCR halt / rebirth | ✅ Stable |
|
||
| Clarification fallback | ✅ Basic |
|
||
| User‑visible “I don’t know” | ✅ Active |
|
||
|
||
---
|
||
|
||
## 📝 Tips & Limits
|
||
|
||
* Works without retriever—manual paste triggers the same checks.
|
||
* Extreme knowledge gaps produce a halt; add sources to continue.
|
||
* Share tricky bluff cases in **Discussions**; they refine ΔS thresholds.
|
||
|
||
---
|
||
|
||
### 🔗 Quick‑Start Downloads (60 sec)
|
||
|
||
| Tool | Link | 3‑Step Setup |
|
||
| -------------------------- | --------------------------------------------------- | ---------------------------------------------------------------------------------------- |
|
||
| **WFGY 1.0 PDF** | [Engine Paper](https://zenodo.org/records/15630969) | 1️⃣ Download · 2️⃣ Upload to LLM · 3️⃣ Ask “Answer using WFGY + \<your question>” |
|
||
| **TXT OS (plain‑text OS)** | [TXTOS.txt](https://zenodo.org/records/15788557) | 1️⃣ Download · 2️⃣ Paste into any LLM chat · 3️⃣ Type “hello world” — OS boots instantly |
|
||
|
||
---
|
||
|
||
> **Help wipe bluffing off the AI map:** drop a ⭐ if this page saved you time.
|
||
> ↩︎ [Back to Problem Index](./README.md)
|