qwed 2.4.0

The Deterministic Verification Protocol for AI - 8 verification engines for math, logic, code, SQL, facts, images, and more.

QWED Protocol

Model Agnostic Verification Layer for AI

QWED Verification - Production-grade deterministic verification layer for Large Language Models. Works with ANY LLM - OpenAI, Anthropic, Gemini, Llama (via Ollama), or any local model. Detect and prevent AI hallucinations through 8 specialized verification engines. Your LLM, Your Choice, Our Verification.

Don't fix the liar. Verify the lie.
QWED does not reduce hallucinations. It makes them irrelevant.

If an AI output cannot be proven, QWED will not allow it into production.

🌐 Model Agnostic: Local ($0) • Budget ($5/mo) • Premium ($100/mo) - You choose!

💖 Support QWED Development:

Quick Start · 🆕 QWEDLocal · The Problem · The 8 Engines · 🔌 Integration · ⚡ QWEDLocal · 🖥️ CLI · 🆓 Ollama (FREE!) · 📖 Full Documentation

---

⚠️ What QWED Is (and Isn't)

QWED is: An open-source engineering tool that combines existing verification libraries (SymPy, Z3, SQLGlot, AST) into a unified API for LLM output validation.

QWED is NOT: Novel research. We don't claim algorithmic innovation. We claim practical integration for production use cases.

Works when: Developer provides ground truth (expected values, schemas, contracts) and LLM generates structured output.

Doesn't work when: Specs come from natural language, outputs are freeform text, or verification domain is unsupported.

🔬 On "Deterministic" Verification

QWED uses deterministic computation (no neural networks, no embeddings, no vibes) wherever possible. Math, Logic, SQL, Code, and Schema engines produce 100% reproducible results using symbolic solvers. For fact-checking, we use TF-IDF (not embeddings) because it's transparent and inspectable—same query always returns same score. For image/reasoning domains that require LLM fallback, we clearly mark outputs as HEURISTIC in the response.

---

🚀 Quick Start: Install & Verify in 30 Seconds

Python SDK (PyPI)

pip install qwed

Go SDK

go get github.com/QWED-AI/qwed-verification/sdk-go

TypeScript SDK (npm)

npm install @qwed-ai/sdk

From Source

git clone https://github.com/QWED-AI/qwed-verification.git
cd qwed-verification
pip install -e .

---

🎓 NEW: Free Course on AI Verification

Learning Path: From Zero to Production-Ready AI Verification

🚀 Start the Free Course →

What You'll Learn:

• 💡 Artist vs. Accountant: Why LLMs are creative but terrible at math
• 🧮 Neurosymbolic AI: How deterministic verification catches 100% of errors*
• 🏗️ Production Patterns: Build guardrails that actually work
• 🔒 HIPAA/GDPR Compliance: PII masking for regulated industries
• 🦜 Framework Integration: LangChain, LlamaIndex, and more

Total Time: ~3 hours | Modules: 4 | Examples: Production-ready code

Perfect for: Developers integrating LLMs, ML engineers, Tech leads evaluating AI safety

---

from qwed_sdk import QWEDClient

client = QWEDClient(api_key="your_key")

# The LLM says: "Derivative of x^2 is 3x" (Hallucination!)
response = client.verify_math(
    query="What is the derivative of x^2?",
    llm_output="3x" 
)

print(response)
# -> ❌ CORRECTED: The derivative is 2x. (Verified by SymPy)

💡 Want to use QWED locally without our backend? Check out QWEDLocal - works with Ollama (FREE), OpenAI, Anthropic, or any LLM provider.

---

🚨 The LLM Hallucination Problem: Why AI Can't Be Trusted

Everyone is trying to fix AI hallucinations by Fine-Tuning (teaching it more data).

This is like forcing a student to memorize 1,000,000 math problems.

What happens when they see the 1,000,001st problem? They guess.

---

📊 The Proof: Why Enterprise AI Needs QWED Verification

We benchmarked Claude Opus 4.5 (one of the world's best LLMs) on 215 critical tasks.

Finding	Implication
Finance: 73% accuracy	Banks can't use raw LLM for calculations
Adversarial: 85% accuracy	LLMs fall for authority bias tricks
QWED: 100% error detection	All 22 errors caught before production

QWED doesn't compete with LLMs. We ENABLE them for production use.

📄 Full Benchmark Report →

---

🎯 Use Cases & Applications

QWED is designed for industries where AI errors have real consequences:

Industry	Use Case	Risk Without QWED
🏦 Financial Services	Transaction validation, fraud detection	$12,889 error per miscalculation
🏥 Healthcare AI	Drug interaction checking, diagnosis verification	Patient safety risks
⚖️ Legal Tech	Contract analysis, compliance checking	Regulatory violations
📚 Educational AI	AI tutoring, assessment systems	Misinformation to students
🏭 Manufacturing	Process control, quality assurance	Production defects

---

✅ The Solution: Verification Layer

QWED is the first open-source Neurosymbolic AI Verification Layer.

We combine:

• Neural Networks (LLMs) for natural language understanding
• Symbolic Reasoning (SymPy, Z3, AST) for deterministic verification

The Core Philosophy: "The Untrusted Translator"

QWED operates on a strict principle: Don't trust the LLM to compute or judge; trust it only to translate.

Example Flow:

User Query: "If all A are B, and x is A, is x B?"

↓ (LLM translates)

Z3 DSL: Implies(A(x), B(x))

↓ (Z3 proves)

Result: TRUE (Proven by formal logic)

The LLM is an Untrusted Translator. The Symbolic Engine is the Trusted Verifier.

---

💡 How QWED Compares

The "Judge" Problem

Most AI safety tools use "LLM-as-a-Judge" (asking GPT-4 to grade GPT-3.5). This is fundamentally unsafe:

• Recursive Hallucination: If the judge has the same bias as the generator, errors go undetected
• Probabilistic Evaluation: LLMs give probability, not proof
• Subjectivity: Different judges = different answers

QWED introduces "Solver-as-a-Judge": Replace neural network opinions with compiler execution and mathematical proof.

Comparison Table

Feature	QWED Protocol	BEAVER	NeMo Guardrails	LangChain Evaluators
The "Judge"	Deterministic Solver (Z3/SymPy)	Probability Bounds	Semantic Matcher	Another LLM (GPT-4)
Mechanism	Translation to DSL	Token Trie Search	Vector Similarity	Prompt Engineering
Verification Type	Mathematical Proof	Probabilistic Bounds	Policy Adherence	Consensus/Opinion
Primary Goal	Correctness (Truth)	Constraint Satisfaction	Safety (Appropriateness)	Quality Score
False Positives	Near Zero (Logic-based)	Low (Statistical)	Medium (Semantic drift)	High (Subjectivity)
Works Offline	✅ Yes (QWEDLocal)	✅ Yes	❌ No	❌ No
Privacy	✅ 100% Local	✅ Local	❌ Cloud-based	❌ Cloud-based

QWED's Advantage: When you need proof, not opinion.

---

🔬 The Verification Engines

QWED routes queries to specialized engines that act as DSL interpreters:

┌──────────────┐
│  User Query  │
└──────┬───────┘
       │
       ▼
┌──────────────────┐
│ LLM (The Guesser)│
│ GPT-4 / Claude   │
└──────┬───────────┘
       │ Unverified Output
       ▼
┌────────────────────┐
│  QWED Protocol     │
│  (Verification)    │
└──────┬─────────────┘
       │
   ┌───┴────┐
   ▼        ▼
❌ Reject  ✅ Verified
            │
            ▼
   ┌────────────────┐
   │ Your Application│
   └────────────────┘

---

QWED 🆚 Traditional AI Safety Approaches

Approach	Accuracy	Deterministic	Explainable	Best For
QWED Verification	✅ 99%+	✅ Yes	✅ Full trace	Production AI
Fine-tuning / RLHF	⚠️ ~85%	❌ No	❌ Black box	General improvement
RAG (Retrieval)	⚠️ ~80%	❌ No	⚠️ Limited	Knowledge grounding
Prompt Engineering	⚠️ ~70%	❌ No	⚠️ Limited	Quick fixes
Guardrails	⚠️ Variable	❌ No	⚠️ Reactive	Content filtering

QWED doesn't replace these - it complements them with mathematical certainty.

---

🔧 The 8 Verification Engines: How QWED Validates LLM Outputs

We don't use another LLM to check your LLM. That's circular logic.

We use Hard Engineering:

Engine	Tech Stack	What it Solves
🧮 Math Verifier	SymPy + NumPy	Calculus, Linear Algebra, Finance. No more $1 + $1 = $3.
⚖️ Logic Verifier	Z3 Prover	Formal Verification. Checks for logical contradictions.
🛡️ Code Security	AST + Semgrep	Catches eval(), secrets, vulnerabilities before code runs.
📊 Stats Engine	Pandas + Wasm	Sandboxed execution for trusted data analysis.
🗄️ SQL Validator	SQLGlot	Prevents Injection & validates schema.
🔍 Fact Checker	TF-IDF + NLI	Checks grounding against source docs.
👁️ Image Verifier	OpenCV + Metadata	Verifies image dimensions, format, pixel data.
🤝 Consensus Engine	Multi-Provider	Cross-checks GPT-4 vs Claude vs Gemini.

---

🧠 The QWED Philosophy: Verification Over Correction

❌ Wrong Approach	✅ QWED Approach
"Let's fine-tune the model to be more accurate"	"Let's verify the output with math"
"Trust the AI's confidence score"	"Trust the symbolic proof"
"Add more training data"	"Add a verification layer"
"Hope it doesn't hallucinate"	"Catch hallucinations deterministically"

QWED = Query with Evidence and Determinism

Probabilistic systems should not be trusted with deterministic tasks. If it can't be verified, it doesn't ship.

---

🔌 LLM Framework Integrations

Already using an Agent framework? QWED drops right in.

🦜 LangChain (Native Integration)

Install: pip install 'qwed[langchain]'

from qwed_sdk.integrations.langchain import QWEDTool
from langchain.agents import initialize_agent
from langchain_openai import ChatOpenAI

# Initialize QWED verification tool
tool = QWEDTool(provider="openai", model="gpt-4o-mini")

# Add to your agent
llm = ChatOpenAI()
agent = initialize_agent(tools=[tool], llm=llm)

# Agent automatically uses QWED for verification
agent.run("Verify: what is the derivative of x^2?")

🤖 CrewAI

from qwed_sdk.integrations.crewai import QWEDVerifiedAgent

agent = QWEDVerifiedAgent(role="Analyst", verify_math=True)

🦙 LlamaIndex

from qwed_sdk.integrations.llamaindex import QWEDQueryEngine

# Add Fact Guard verification to any query engine
verified_engine = QWEDQueryEngine(base_engine, verify_facts=True)

---

🛒 QWED Integrations & Extensions

QWED extends beyond core verification with specialized integrations:

Integration	Description	Repo
🛒 QWED-UCP	Unified Commerce Protocol - Verify e-commerce AI (prices, inventory, transactions)
🤖 Open Responses	OpenAI Responses API integration with QWED verification guards
🔌 QWED-MCP	Model Context Protocol server - Use QWED as Claude Desktop tool

🛒 QWED-UCP (Commerce Verification)

Prevent AI hallucinations in e-commerce: wrong prices, fake inventory, invalid transactions.

from qwed_ucp import UCPGuard

guard = UCPGuard()
result = guard.verify_transaction({
    "product": "iPhone 15",
    "price": 999.00,
    "quantity": 2,
    "total": 1998.00  # AI calculated this
})
# ✅ VERIFIED: total = price × quantity

🔌 QWED-MCP (Claude Desktop)

Use QWED verification directly in Claude Desktop via MCP:

{
  "mcpServers": {
    "qwed": {
      "command": "uvx",
      "args": ["qwed-mcp"]
    }
  }
}

---

🌍 Multi-Language SDK Support

Language	Package	Status
🐍 Python	qwed	✅ Available on PyPI
🟦 TypeScript	@qwed-ai/sdk	✅ Available on npm
🐹 Go	qwed-go	✅ Available
🦀 Rust	qwed	🟡 Coming Soon

# Python
pip install qwed

# Go
go get github.com/QWED-AI/qwed-verification/sdk-go

# TypeScript
npm install @qwed-ai/sdk

---

🎯 Real Example: The $12,889 Bug

User asks AI: "Calculate compound interest: $100K at 5% for 10 years"

GPT-4 responds: "$150,000"
(Used simple interest by mistake)

With QWED:

response = client.verify_math(
    query="Compound interest: $100K, 5%, 10 years",
    llm_output="$150,000"
)
# -> ❌ INCORRECT: Expected $162,889.46
#    Error: Used simple interest formula instead of compound

Cost of not verifying: $12,889 error per transaction 💸

---

❓ Frequently Asked Questions

Q: How does QWED differ from RAG (Retrieval Augmented Generation)?

A: RAG improves the input to the LLM by grounding it in documents. QWED verifies the output deterministically. RAG adds knowledge; QWED adds certainty.

Q: Can QWED work with any LLM?

A: Yes! QWED is model-agnostic and works with GPT-4, Claude, Gemini, Llama, Mistral, and any other LLM. We verify outputs, not models.

Q: Does QWED replace fine-tuning?

A: No. Fine-tuning makes models better at tasks. QWED verifies they got it right. Use both.

Q: Is QWED open source?

A: Yes! Apache 2.0 license. Enterprise features (audit logs, multi-tenancy) are in a separate repo.

Q: What's the latency overhead?

A: Typically <100ms for most verifications. Math and logic proofs are instant. Consensus checks take longer (multiple API calls).

---

📚 Documentation & Resources

Main Documentation:

Resource	Description
📖 Full Documentation	Complete API reference and guides
🔧 API Reference	Endpoints and schemas
⚡ QWEDLocal Guide	Client-side verification setup
🖥️ CLI Reference	Command-line interface
🔒 PII Masking Guide	HIPAA/GDPR compliance
🆓 Ollama Integration	Free local LLM setup

Project Documentation:

Resource	Description
📊 Benchmarks	LLM accuracy testing results
🗺️ Project Roadmap	Future features and timeline
📋 Changelog	Version history summary
📜 Release Notes	Detailed version release notes
🎬 GitHub Action Guide	CI/CD integration
🏗️ Architecture	System design and engine internals

Community:

Resource	Description
🤝 Contributing Guide	How to contribute to QWED
📜 Code of Conduct	Community guidelines
🔒 Security Policy	Reporting vulnerabilities
📖 Citation	Academic citation format

---

🏢 Enterprise Features

Need observability, multi-tenancy, audit logs, or compliance exports?

📧 Contact: [e-mail]

---

📄 License

Apache 2.0 - See LICENSE

---

⭐ Star History

If chart doesn't load, click here for alternatives

Current Stars:

View trend: Star History Page

---

👥 Contributors

---

📄 Citation

If you use QWED in your research or project, please cite our archived paper:

@software{dass2025qwed,
  author = {Dass, Rahul},
  title = {QWED Protocol: Deterministic Verification for Large Language Models},
  year = {2025},
  publisher = {Zenodo},
  version = {v1.0.0},
  doi = {10.5281/zenodo.18110785},
  url = {https://doi.org/10.5281/zenodo.18110785}
}

Plain text:

Dass, R. (2025). QWED Protocol: Deterministic Verification for Large Language Models (Version v1.1.0). Zenodo. https://doi.org/10.5281/zenodo.18110785

---

✅ Using QWED in Your Project?

Add this badge to your README to show you're using verified AI:

[![Verified by QWED](https://img.shields.io/badge/Verified_by-QWED-00C853?style=flat&logo=checkmarx)](https://github.com/QWED-AI/qwed-verification#%EF%B8%8F-what-does-verified-by-qwed-mean)

Preview:

This badge tells users that your LLM outputs are deterministically verified, not just "hallucination-prone guesses."

🛡️ What does "Verified by QWED" mean?

When you see the [Verified by QWED] badge on a repository or application, it is a technical guarantee, not a marketing claim.

It certifies that the software adheres to the QWED Protocol for AI Safety:

1. The Zero-Hallucination Warranty: The application does not rely on LLM probabilities for Math, Logic, or Code. It uses Deterministic Engines (SymPy, Z3, AST) to prove correctness before outputting data.

2. The "Untrusted Translator" Architecture: The system treats the LLM solely as a translator (Natural Language → DSL), never as a judge. If the translation cannot be mathematically proven, the system refuses to answer rather than guessing.

3. Cryptographic Accountability: The application generates JWT-based Attestations (ES256 signatures) for its critical operations. Every "Verified" output comes with a cryptographic receipt proving a solver validated it.

In short: The badge means "We don't trust the AI. We trust the Math."

---

⭐ Star us if you believe AI needs verification

Ready to trust your AI?

"Safe AI is the only AI that scales."

Contribute · Architecture · Security · Documentation