🛡️ MuSAW — Cross-Model AI Output Verifier

The independent truth layer for any AI model. Extract claims, detect hallucinations, and find real evidence — works on Claude, GPT, Perplexity, Gemini, Llama, any model.

npx musaw verify output.txt

The Gap

| Company | Can verify itself? | Why not? | |---------|:------------------:|----------| | Anthropic | ❌ | Claude can't audit Claude — same blind spots | | OpenAI | ❌ | GPT shares the same hallucination patterns | | Perplexity | ❌ | Their search summaries are the product itself | | Google (Gemini) | ❌ | Same architecture, same weaknesses | | MuSAW | ✅ | Model-agnostic, no shared DNA |

Every major AI lab is racing to build self-verification. None can succeed. True verification requires independence — a layer that doesn't share the model's training data, architecture, or biases.

Install

# Install globally
npm install -g musaw

# Or run directly
npx musaw

# Or use as a library
npm install musaw

Quick Start

# Run the demo
npx musaw demo

# Verify a file
npx musaw check output.txt

# Pipe from another command
cat output.txt | npx musaw check -

# Quick confidence check
npx musaw quick output.txt

# Compare two model outputs
npx musaw compare output-a.txt output-b.txt

# Run tests
npx musaw test

Use as a Library

import { MuSAW } from 'musaw';

const musaw = new MuSAW({ strictness: 'strict' });

// Verify any AI output
const result = await musaw.analyze(`
  According to a recent study published in Nature, 
  Google's Sycamore processor achieved quantum supremacy in 2019.
`);

console.log(result.report);
// ✅ Found: "Sodium Regulation in the Blood..." (Nature, 1959)
//    DOI: 10.1038/184283a0
//    Confidence: 60% — source exists but claim not verified

// Import individual modules
import { ClaimExtractor } from 'musaw/core/claim-extractor';
import { Verifier } from 'musaw/core/verifier';
import { EvidenceRetriever } from 'musaw/core/evidence-retriever';
import { ReportGenerator } from 'musaw/core/report-generator';

Architecture

AI Output ──→ MuSAW ──→ Per-Claim Confidence Score + Evidence
(any model)

Layer 1: CLAIM EXTRACTION
  Parses AI output into atomic verifiable statements.
  Identifies: entities, numbers, dates, sources, forgeability patterns
  Flags: "studies show", "experts say", over-precise percentages

Layer 2: VERIFICATION ENGINE (5 dimensions)
  • Internal consistency — does it contradict itself?
  • Numeric sanity — do the numbers make sense?
  • Attribution quality — is the source real and specific?
  • Temporal consistency — are timeframes realistic?
  • Hallucination patterns — known fabrication signatures?

Layer 3: EVIDENCE RETRIEVAL
  For claims referencing studies/journals:
  • Extracts DOIs, journal names, paper titles
  • Queries CrossRef API (free, no key needed)
  • Returns: real paper title, DOI, journal, author, direct URL
  • Boosts confidence when real source confirms the claim

What MuSAW Detects

MuSAW automatically flags:

• "Studies show" / "Research indicates" — unattributed research claims
• "Experts say" / "Some believe" — vague expert consensus
• Over-precise statistics — "exactly 83.7%" is often fabricated
• Fake paper IDs — "paper-2024-xyz" pattern
• Self-contradictions — saying "all" and "some" in the same breath
• Temporal vagueness — "recently", "soon", "historically"
• Suspiciously round numbers — 42, 100, 1000 popping up

Verified Models

MuSAW works on any text output from any model, with zero changes:

| Model | Company | Status | |-------|---------|:------:| | Claude | Anthropic | ✅ | | GPT-4 / GPT-4o | OpenAI | ✅ | | Perplexity | Perplexity AI | ✅ | | Gemini | Google | ✅ | | Llama | Meta | ✅ | | Mistral | Mistral AI | ✅ | | Grok | xAI | ✅ | | DeepSeek | DeepSeek | ✅ | | Copilot | Microsoft | ✅ | | Any future model | Any company | ✅ |

Project Structure

musaw/
├── src/
│   ├── cli.js                    # CLI entry with demo/test/compare
│   ├── pitch.js                  # $100M pitch deck
│   ├── targets.js                # 14 acquisition targets
│   ├── verify-all-models.js      # Cross-model demo
│   └── core/
│       ├── musaw.js              # Main orchestrator
│       ├── claim-extractor.js    # Text → atomic claims
│       ├── verifier.js           # 5-dimension verification
│       ├── evidence-retriever.js # CrossRef real paper lookup
│       └── report-generator.js   # Beautiful reports
├── workspace/                    # Analysis outputs
├── vault/                        # Trusted fact storage (future)
├── package.json
├── README.md
└── LICENSE

Why This Matters

The industry hallucination rate is 10-40% across all major models. That's $100B+ in liability. The EU AI Act makes verification mandatory by 2026. Every enterprise deploying AI needs an independent verifier.

MuSAW is the only one.

License

MIT — MuSAW Labs

Pitch

npx musaw pitch   # View $100M pitch deck
npx musaw targets  # View acquisition targets