Real-time, government-backed data fact-checking for a better, more transparent internet.
November 19, 2025 - Project Mirador Chrome extension is now live on the Chrome Web Store!
Install: Search for "Project Mirador" on the Chrome Web Store or use the link above.
- About
- Key Features
- How It Works
- Getting Started
- Tech Stack
- Architecture
- API Documentation
- Contributing
- Disclaimer
- License
- Contact
This repository contains the codebase for Project Mirador and the Stelthar API.
Stelthar API provides real-time, government-backed data fact-checking for a better, more transparent version of the internet.
Disinformation erodes public trust, and traditional fact-checking is slow, unreliable, or both.
Stelthar's fact-checking API verifies claims by comparing them against official government data sources. Highlight any text, and our browser extension provides an evidence-based verdict, a confidence score, and links to the primary sources.
This reduces verification time from hours to seconds, empowering users to challenge misinformation with auditable proof.
- Real-time claim verification against official government data sources
- Evidence-based verdicts: Supported, Contradicted, or Inconclusive
- Confidence scoring that reflects data coverage and match quality
- Direct links to the primary sources used for verification
- Browser extension integration for on-page highlight-and-check functionality
- API-first design for easy integration into apps, dashboards, and extensions
- Highlight: Select any text claim on a webpage
- Verify: Click the Project Mirador extension icon
- Review: Receive an instant verdict with:
- Supported/Contradicted/Inconclusive status
- Confidence score (0-100%)
- Links to official government sources
- Relevant data points used for verification
Under the Hood:
- Your claim is sent to the Stelthar API
- The API queries multiple government databases (Data.gov, BLS, BEA, etc.)
- AI-powered analysis compares your claim against official data
- Results are returned with full source attribution
- Visit the Chrome Web Store
- Click "Add to Chrome"
- Pin the extension to your toolbar for easy access
- Navigate to any webpage and highlight text to fact-check
- Click the Project Mirador icon to verify the claim
The Stelthar API powers Project Mirador and is available for integration into your own applications.
Quick Example:
// Example API request (pseudocode)
fetch('https://api.stelthar.com/verify', {
method: 'POST',
headers: {
'Content-Type': 'application/json',
},
body: JSON.stringify({
claim: "The unemployment rate in 2024 was 3.7%"
})
})
.then(response => response.json())
.then(data => console.log(data));See the API Documentation section below for detailed endpoint information.
- Language: Python
- Hosting: Vercel (API hosting)
- Languages: JavaScript, HTML, CSS
- Platform: Chrome Extension
Structured (return actual numbers — always preferred):
- USAspending.gov — federal agency budgets & obligations by fiscal year (no API key)
- Treasury Fiscal Data — national debt / public debt outstanding (no API key)
- Bureau of Economic Analysis (BEA) — GDP, personal income, government receipts & expenditures by function, trade (12 NIPA tables)
- Bureau of Labor Statistics (BLS) — unemployment, CPI, core CPI, PPI, labor force participation, employment level, average hourly earnings
- U.S. Census Bureau (ACS) — population, income, poverty, age, home value, education, insurance coverage, by state or nationally
Keyword search (return dataset descriptions, not values — supplement only):
- Data.gov catalog
- Congress.gov (bills & legislation)
- Gemini API (restrained for fact-checking purposes)
- API Testing: Postman
- Version Control: Git & GitHub
Stelthar runs an agentic (multi-pass) evidence pipeline. Rather than issuing one query plan and giving up when it misses, the engine critiques what it retrieved and re-plans against the specific gap.
- Plan: An LLM routes the claim to the source that can actually answer it (agency budget → USAspending, national debt → Treasury, GDP → BEA, inflation → BLS, demographics → Census with server-side variable resolution).
- Retrieve: All selected sources are queried in parallel, deduplicated across passes.
- Critique: A dedicated critic checks whether the evidence contains a real datapoint addressing the claim's entities and timeframe. Dataset catalog pages are explicitly rejected as non-evidence.
- Re-plan: If evidence is insufficient, the critic proposes a corrected plan and the
loop retries (bounded by
MAX_ITERATIONS). - Synthesize: The LLM reasons only over ranked evidence, with valueless sources labelled so they cannot be mistaken for proof.
- Score: Confidence is capped when evidence doesn't earn it (no datapoint, low relevance, or year mismatch).
┌─────────────────┐
│ User Browser │
│ (Chrome Ext) │
└────────┬────────┘
│ Highlighted OR Pasted Claim
▼
┌──────────────────────────┐
│ Claim Analysis (LLM) │──► routes to the source that has the answer
└────────┬─────────────────┘
▼
┌──────────────────────────────────────────────┐
│ AGENTIC LOOP (≤3 passes) │
│ │
│ ┌────────────┐ ┌───────────────────┐ │
│ │ Retrieve │─────►│ Critique │ │
│ │ (parallel)│ │ • real datapoint?│ │
│ └─────▲──────┘ │ • right year? │ │
│ │ │ • on topic? │ │
│ │ └─────────┬─────────┘ │
│ │ insufficient │ sufficient │
│ └──── re-plan ◄─────────┤ │
└──────────────────────────────────────────────┘
│
▼
┌──────────────────────┐
│ Rank evidence │ numbers first, catalog pages last
│ Synthesize (LLM) │
│ Score + cap │
└──────┬───────────────┘
▼
┌──────────────────────────────┐
│ Verdict + Confidence │
│ + Sources + debug_iterations │
└──────────────────────────────┘
https://api.stelthar.com
Verify a claim against government data sources.
Request Body:
{
"claim": "The United States federal government spent more on defense than on education in 2023."
}Response:
{
"claim_original": "The United States federal government spent more on defense than on education in 2023.",
"claim_normalized": "US federal spending: defense > education (2023)",
"claim_type": "quantitative_comparison",
"verdict": "Supported",
"confidence": 0.75,
"confidence_tier": "Medium",
"confidence_breakdown": {
"source_reliability": 0.7,
"evidence_density": 1.0,
"semantic_alignment": 0.66
},
"summary": "Federal defense spending ($790.9B) exceeded education spending ($178.6B) in 2023 according to BEA NIPA data.",
"evidence_links": [
{
"finding": "Defense: $790.9B",
"source_url": "https://apps.bea.gov/NIPA/T31600"
},
{
"finding": "Education: $178.6B",
"source_url": "https://apps.bea.gov/NIPA/T31600"
}
],
"sources": [
{
"title": "BEA NIPA T31600 - National Defense",
"url": "https://apps.bea.gov/NIPA/T31600",
"snippet": "2023 defense spending: $790.2B",
"data_value": 790197.0,
"unit": "Millions USD"
},
{
"title": "BEA NIPA T31600 - Education",
"url": "https://apps.bea.gov/NIPA/T31600",
"snippet": "2023 education spending: $178.6B",
"data_value": 178621.0,
"unit": "Millions USD"
}
],
"debug_plan": {
"claim_type": "quantitative_comparison",
"entities": ["defense", "education", "2023"],
"api_plan": {
"tier1": { "bea": { "table": "T31600", "year": "2023", "lines": ["2", "14"] } },
"tier2_keywords": ["federal spending 2023"]
}
},
"debug_log": []
}Rate Limits: 100 requests/day
Any contributions you make are greatly appreciated.
-
Clone the repository
git clone https://github.com/RadoKyselak/Project-Mirador.git cd Project-Mirador -
Install dependencies
# Backend dependencies pip install -r requirements.txt # Frontend - load the extension in Chrome # Navigate to chrome://extensions/ # Enable "Developer mode" # Click "Load unpacked" and select the extension directory
-
Set up environment variables
cp .env.example .env # Add your API keys for government data sources and Gemini -
Run tests
pytest tests/
- Fork the repository
- Create a feature branch (
git checkout -b feature/AmazingFeature) - Commit your changes (
git commit -m 'Add some AmazingFeature') - Push to the branch (
git push origin feature/AmazingFeature) - Open a Pull Request
Stelthar API uses publicly available U.S. Government open data sources but is not affiliated with or endorsed by any governmental agency.
Verdicts and confidence scores are algorithmic estimates — they assist, not replace, human judgment. Always verify critical information through multiple sources.
This project is licensed under the MIT License - see the LICENSE file for details.
Rado Kyselak - Developer
- GitHub: @RadoKyselak
Open Truth. Verified Data.
Project Mirador - Powered by Stelthar API