Your AI is making things up. You just don't know when.
Stop AI from lying in production
Wauldo Guard is a hallucination firewall that blocks wrong answers before they reach your users.
# Your existing code
response = openai.chat.completions.create(...)
# Add Wauldo Guard
verified = wauldo.fact_check(response, source="your_document")
if verified.verdict == "verified":
return response # safe to show
else:
return "I don't have enough evidence to answer."
The problem no one tells you
Every LLM sounds confident. Every LLM generates plausible answers. And every LLM still gets things wrong. If your AI is live, it has already hallucinated. Your users have already seen it.
What happens without a guard
- × Wrong answers reach users
- × Trust drops instantly
- × Legal and compliance risks
- × Silent failures you can't detect
LLMs don't fail loudly. They fail convincingly.
See it in action
Without Guard
"Returns are accepted within 60 days of purchase."
Sounds right. Looks confident. But the actual policy says 14 days.
Your user just got wrong information.
With Wauldo Guard
verdict: "rejected"
action: "block"
reason: "numerical_mismatch"
confidence: 0.03
Caught. Blocked. User protected.
Most RAG APIs
Retrieve → Generate → Hope
Wauldo Guard
The fix: add a firewall
One API call between your LLM and your users. If it's wrong, it never ships.
- ✓ Validate claims against source documents
- ✓ Detect numerical mismatches and contradictions
- ✓ Check citation coverage (phantom references)
- ✓ Score confidence (0-1)
- ✓ Block or flag unsafe outputs
Not another AI tool
Guard doesn't generate answers. It catches the bad ones.
- × Not a chatbot
- × Not a RAG system
- × Not a wrapper
- ✓ A safety layer for AI systems in production
We don't improve AI. We stop it from being wrong.
Your AI can now say "I don't know" instead of guessing. That's the difference between a demo and a product.
Three verification modes
Choose speed or depth. Same API.
Lexical
<1ms
Word overlap + contradiction detection. No model needed.
Hybrid
~50ms
Keyword + BGE semantic similarity. Catches paraphrases.
Semantic
~500ms
Full embeddings. "Chapter 11" = "went bankrupt".
Citation
<1ms
Detect phantom references and uncited claims.
If your AI talks to users, you need this
No exception.
Customer support AI
Your bot stops inventing features that don't exist. Wrong refund policy = real money lost.
Legal & compliance
Every AI output verified against policy before reaching the user. Audit trail included.
Internal copilots
Employees get answers from company docs, not LLM imagination. No more "the AI said so" incidents.
AI-powered SaaS
Ship AI features without shipping hallucinations. Your users trust you — don't break that.
Works with any LLM
OpenAI, Claude, Gemini, Llama, Mistral — doesn't matter. Guard checks the output, not the model.
# Fact-check a claim against source
result = wauldo.fact_check(
text="Returns accepted within 60 days",
source_context="Policy allows returns within 14 days",
mode="lexical"
)
result.verdict # "rejected"
result.action # "block"
result.reason # "numerical_mismatch"
result.confidence # 0.03# Verify citations
result = wauldo.verify_citation(
text="Rust was created in 2010 [Source: docs]. It is fast [Source: fake].",
sources=[{"name": "docs", "content": "Rust released in 2010."}]
)
result.citation_ratio # 1.0
result.phantom_count # 1 (fake source detected)Simple pricing
Start free. Pay when you scale.
Need more? $0.002/check unlimited.
Stop shipping hallucinations
Start blocking wrong answers now. 2 lines of code. Free to start.