GauntletScore — Trust Verification for AI Output

Verify any AI-generated document or code before you trust it. Seven AI personas from five independent model providers independently analyze your content, verify every checkable claim against authoritative sources, and produce a cryptographically signed trust score.

What It Does

Submit a document or code and get:

• Gauntlet Score (0-100) with letter grade (A-F)
• Claim-by-claim verification against CourtListener (legal), eCFR (regulatory), PubMed (scientific), EDGAR (SEC), and computational math verification
• Code safety analysis detecting reverse shells, credential theft, prompt manipulation, data exfiltration, and obfuscated payloads
• Unanimous vote from 7 independent AI agents (PROCEED / PROCEED WITH CONDITIONS / DO NOT PROCEED)
• Cryptographic certificate proving the score is genuine and untampered
• Full debate transcript showing every agent's reasoning

Quick Start

Free tier includes 3 analyses per month. Get an API key at gauntletscore.com.

Verify a document by pasting content:

POST https://api.gauntletscore.com/v1/analyze
Authorization: Bearer YOUR_API_KEY
Content-Type: application/json

{
  "document": "Your document text here...",
  "topic": "Verify the claims in this document"
}

Verify a ClawHub skill by URL:

POST https://api.gauntletscore.com/v1/analyze
Authorization: Bearer YOUR_API_KEY
Content-Type: application/json

{
  "source_url": "https://clawhub.ai/skills/gauntlet-validate/SKILL.md",
  "topic": "Evaluate the safety of this ClawHub skill before installation"
}

Check results:

GET https://api.gauntletscore.com/v1/jobs/{job_id}
Authorization: Bearer YOUR_API_KEY

Results include score, grade, vote, verified claims, and a cryptographic certificate.

What It Catches

In documents:

• Fabricated legal citations (hallucinated case law)
• Misapplied regulations (wrong CFR section for the situation)
• Mathematical errors (wrong totals, incorrect percentages)
• Internal contradictions
• Unsupported conclusions

In code:

• Reverse shells and remote code execution
• Credential theft and data exfiltration
• Download-and-execute attacks (curl | bash)
• Prompt manipulation attempts (instructions designed to override agent behavior)
• Documentation-vs-behavior mismatches (code does things the README doesn't mention)
• Dangerous operation combinations that pass individual checks

How It Works

Seven AI personas from five independent model providers independently analyze your submission:

1. Round 0 — Each agent conducts independent research, verifying claims against authoritative databases
2. Rounds 1-3 — Structured adversarial debate where agents challenge each other's findings
3. Round 4 — Final positions and votes
4. Knowledge Graph — Every verified and debunked claim is stored in a persistent knowledge graph. Subsequent analyses benefit from prior verifications, reducing cost and increasing accuracy over time.
5. Bayesian Calibration — Confidence scores are computed using Bayesian inference across multiple evidence sources, not simple vote counting. The score reflects calibrated probability, not consensus.
6. Scoring — Six-component rubric produces the Gauntlet Score
7. Certification — Ed25519 cryptographic signature proves the result is genuine

All analysis is read-only. Submitted code is never executed. Documents are processed in memory and not stored.

Pricing

One credit = one analysis, regardless of document length. No subscriptions. See gauntletscore.com/pricing for details. Sovereign Edition: [email protected]

Verify a Certificate

Anyone can verify a Gauntlet Score is genuine:

GET https://api.gauntletscore.com/v1/verify/{certificate_id}

No authentication required.

Links

• API Documentation
• Terms of Service
• Privacy Policy
• Acceptable Use Policy

About

GauntletScore is built by Genstrata, Inc. The Gauntlet's adversarial multi-agent verification architecture is patent-pending (USPTO #63/967,169).

For organizations that cannot send data to cloud services, the Sovereign Edition runs entirely on your hardware with zero data egress. Contact [email protected].