mirror of
https://github.com/onestardao/WFGY.git
synced 2026-05-02 21:41:09 +00:00
180 lines
7.5 KiB
Markdown
180 lines
7.5 KiB
Markdown
<!-- ======================================================= -->
|
||
<!-- agent-consensus-protocols.md · Semantic Clinic / Map-B -->
|
||
<!-- Draft v0.1 · MIT · 2025-08-06 -->
|
||
<!-- Purpose: Design rock-solid agreement loops for 2-N -->
|
||
<!-- agents (debate, Critic-Coder, planning crews) while -->
|
||
<!-- keeping speed ⬆, token cost ↓, and hallucination = 0. -->
|
||
<!-- ======================================================= -->
|
||
|
||
# 🤝 Agent Consensus Protocols
|
||
*Make two (or ten) LLMs reach a decision **on time** and **on budget** — every run.*
|
||
|
||
> **Why this page?**
|
||
> “Let them debate” sounds cool until:
|
||
> * the Critic never stops nit-picking,
|
||
> * the Planner forgets half the context,
|
||
> * the Coder ships five conflicting drafts, and
|
||
> * your bill looks like a crypto rug-pull.
|
||
>
|
||
> WFGY adds measurable **semantic tension (ΔS)** and **logic vectors (λ)** to debate & QA loops, so you can **prove** consensus is reached — not guessed.
|
||
|
||
---
|
||
|
||
## 1 · Three Canonical Consensus Modes
|
||
|
||
| Mode | Use Case | # Agents | Target ΔS | Target λ |
|
||
|------|----------|---------:|----------:|----------|
|
||
| **Critic-Coder** | Code/SQL generation with review | 2 | ≤ 0.40 | convergent |
|
||
| **Triad Debate** | Fact-check or legal reasoning | 3 | ≤ 0.45 | convergent |
|
||
| **Crew Vote** | 4 – 10 micro-agents on long plan | N | ≤ 0.50 | majority convergent |
|
||
|
||
*ΔS ceiling rises gently with group size; λ must return to convergent after each round.*
|
||
|
||
---
|
||
|
||
## 2 · Failure Modes (Why Votes Stall)
|
||
|
||
| # | Failure | ΔS / λ Symptom | Example |
|
||
|--:|---------|----------------|---------|
|
||
| 1 | **Infinite Debate** | λ oscillates (→, ←, →…) | Critic re-opens same defect |
|
||
| 2 | **Topic Drift** | ΔS > 0.60 vs. original prompt | Agents add new, unrelated goals |
|
||
| 3 | **Split Brain** | Two stable but divergent λ | 50 / 50 vote, no tie-breaker |
|
||
| 4 | **Early Merge** | ΔS still 0.55 but λ forced convergent | “Agree to disagree” w/ wrong answer |
|
||
| 5 | **Token Exhaust** | E_resonance ↑ while ΔS unchanged | Loop burns context without progress |
|
||
|
||
---
|
||
|
||
## 3 · WFGY Consensus Blueprint
|
||
|
||
> Four guard layers; plug into any framework (AutoGen, LangChain, custom asyncio).
|
||
|
||
| Layer | Module | Guard | Effect |
|
||
|-------|--------|-------|--------|
|
||
| **Round ΔS Checker** | BBMC | ΔS(agent, goal) ≤ ceiling | Drop off-topic vote |
|
||
| **λ Mediator** | ΔS + λ | λ must converge within 3 rounds | Force re-focus or escalate |
|
||
| **Vote Auditor** | WAI | JSON vote schema + arg defaults | No partial / malformed votes |
|
||
| **BBCR Fallback** | BBCR | > 5 rounds OR ΔS stagnates | Collapse & request human tie-break |
|
||
|
||
```mermaid
|
||
flowchart TD
|
||
subgraph Round
|
||
Q[User Goal]
|
||
A1[Agent 1]
|
||
A2[Agent 2]
|
||
A3[Agent 3]
|
||
Q --> A1 --> V1[Vote JSON]
|
||
Q --> A2 --> V2
|
||
Q --> A3 --> V3
|
||
end
|
||
BBMC -->|ΔS filter| V1
|
||
BBMC --> V2
|
||
BBMC --> V3
|
||
V1 & V2 & V3 --> Mediator[λ Mediator] -->|convergent?| Decision
|
||
Mediator -.->|no| BBCR
|
||
````
|
||
|
||
---
|
||
|
||
## 4 · Practical Setup (AutoGen example)
|
||
|
||
```python
|
||
from autogen import AssistantAgent, UserProxyAgent
|
||
from wfgy import consensus_filter
|
||
|
||
critic = AssistantAgent(name="critic")
|
||
coder = AssistantAgent(name="coder")
|
||
user = UserProxyAgent("user")
|
||
|
||
def callback(messages, state):
|
||
# 1. measure ΔS + λ
|
||
ok = consensus_filter(messages, ceiling=0.45, rounds=3)
|
||
if not ok:
|
||
return "STOP_DEBATE"
|
||
return "CONTINUE"
|
||
|
||
coder.register_reply(callback)
|
||
critic.register_reply(callback)
|
||
```
|
||
|
||
*15 lines → loop stops automatically when convergence proven or impossible.*
|
||
|
||
---
|
||
|
||
## 5 · Debug Walk-Through
|
||
|
||
1. **Log votes**
|
||
|
||
```python
|
||
print(state.votes) # {critic: "reject, ΔS=0.48", coder: "accept, ΔS=0.41"}
|
||
```
|
||
|
||
2. **Inspect λ trend**
|
||
|
||
```
|
||
λ sequence: → → ← → (oscillating) ❌
|
||
```
|
||
|
||
> Two divergent rounds trigger Mediator → requests narrowing question.
|
||
|
||
3. **Trigger BBCR**
|
||
|
||
```
|
||
ΔS stagnates at 0.52 for 3 rounds → BBCR => "Need human tie-break"
|
||
```
|
||
|
||
---
|
||
|
||
## 6 · Best-Practice Table
|
||
|
||
| Tip | Why |
|
||
| ------------------------------------------ | ------------------------------------ |
|
||
| **Keep each vote to max 300 tokens.** | Reduces E\_resonance; easier ΔS calc |
|
||
| **Pin goal & constraints in every round.** | Prevents silent prompt drift |
|
||
| **Summarise before vote.** | Normalises embeddings → fair ΔS |
|
||
| **Record vote reason (1-2 lines).** | Faster root-cause when split-brain |
|
||
|
||
---
|
||
|
||
## 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 |
|
||
|
||
---
|
||
|
||
↩︎ [Back to Problem Index](./README.md)
|
||
|
||
---
|
||
|
||
### 🧭 Explore More
|
||
|
||
| Module | Description | Link |
|
||
| --------------------- | ---------------------------------------------------------------- | ----------------------------------------------------------------------------------- |
|
||
| Semantic Blueprint | Layer-based symbolic reasoning & semantic modulations | [View →](https://github.com/onestardao/WFGY/tree/main/SemanticBlueprint) |
|
||
| Benchmark vs GPT-5 | Stress-test GPT-5 with full WFGY reasoning suite | [View →](https://github.com/onestardao/WFGY/tree/main/benchmarks/benchmark-vs-gpt5) |
|
||
| Semantic Clinic Index | Full failure catalog: prompt injection, memory bugs, logic drift | [View →](./SemanticClinicIndex.md) |
|
||
|
||
---
|
||
|
||
> 👑 **Early Stargazers: [Hall of Fame](https://github.com/onestardao/WFGY/tree/main/stargazers)** — engineers & hackers who backed WFGY from day one. <img src="https://img.shields.io/github/stars/onestardao/WFGY?style=social" alt="stars"> **Star the repo** to unlock Engine 2.0 sooner.
|
||
|
||
<div align="center">
|
||
|
||
[](https://github.com/onestardao/WFGY)
|
||
|
||
[](https://github.com/onestardao/WFGY/tree/main/OS)
|
||
|
||
[](https://github.com/onestardao/WFGY/tree/main/OS/BlahBlahBlah)
|
||
|
||
[](https://github.com/onestardao/WFGY/tree/main/OS/BlotBlotBlot)
|
||
|
||
[](https://github.com/onestardao/WFGY/tree/main/OS/BlocBlocBloc)
|
||
|
||
[](https://github.com/onestardao/WFGY/tree/main/OS/BlurBlurBlur)
|
||
|
||
[](https://github.com/onestardao/WFGY/tree/main/OS/BlowBlowBlow)
|
||
|
||
</div>
|
||
|