verifier-mcp

"The model proposes; deterministic executed checks own the verdict."

Deterministic EU e-Invoice compliance engine and paradox resolver. Exposed as a Model Context Protocol (MCP) server for AI coding agents and an instant CLI via npx.

Built from the DevNetwork API + Cloud + AI Hackathon 2026 project.

Why verifier-mcp?

LLM extraction engines hallucinate and report misleading confidence scores on unstructured PDFs. When the EU September 2026 e-invoicing mandate takes effect, automated workflows cannot rely on fuzzy probabilistic guesses.

verifier-mcp wraps document extraction with deterministic verification gates:

1. Contradiction / Paradox Detection: Flags mathematical, temporal, and legal jurisdiction discrepancies before data enters accounting ledgers.
2. 5-Tier Confidence Hierarchy: Routes extractions from Level 1 Straight-Through processing down to Level 5 hard rejection.
3. Single Token Language: Emits an immutable SHA-256 truth token (document_id:resolution:timestamp) as cryptographic audit evidence.

Quick Start via NPX

Run directly without installation:

# Test built-in paradox checks
npx verifier-mcp test-paradox

# Audit a JSON invoice extraction
npx verifier-mcp audit ./invoice.json

Setup as an MCP Server

Add to your AI agent configuration (Claude Desktop, Cursor, Antigravity):

{
  "mcpServers": {
    "verifier": {
      "command": "npx",
      "args": ["-y", "verifier-mcp"]
    }
  }
}

Tools Exposed

| Tool | Parameters | Description | |---|---|---| | detect_paradoxes | fields: object | Detects mathematical mismatches, causality violations, and VAT jurisdiction anomalies. | | evaluate_compliance_tier | confidence: number | Maps confidence scores to the 5-tier EU regulatory review framework. | | generate_truth_token | documentId, resolution, timestamp? | Generates SHA-256 immutable audit proof token. | | audit_invoice | invoiceData: object | Full end-to-end verification audit producing compliance verdict and truth token. | | explain_paradox | paradoxType: string | Returns deterministic mitigation steps for detected contradiction categories. |

License

MIT © Alexander Sorrell