n8n Judgment API: Verify Contract Terms & Flag Hallucinations

Verify contract fields with Judgment—automatically flag mismatches and hallucinations

This n8n workflow uses the Judgment API to verify extracted contract terms against the original contract text. For each field, it confirms what’s supported by the source—or routes suspicious, missing, or conflicting values to a review queue with a human-readable explanation.

What this workflow does

Built for manual testing and evaluation, the workflow runs when you click “Test workflow” and performs a field-by-field consistency check:

• Generates two sample extractions from the same source_text (one accurate, one containing fabricated values).
• Merges both extraction records into a single stream so Judgment can evaluate them side by side.
• Sends each record to Judgment to evaluate, per field:
  • whether the extracted value is missing from the source text
  • whether the source states a different value
• Reshapes results into per-field checks, including field name, check type, extracted value, risk score, and a verdict of confirmed vs needs review.
• Routes results intelligently:
  • If confirmed, the field is treated as verified.
  • If needs review, it adds an explanation and routes the item to a review queue.

Use cases

• Validate outputs from a contract extraction pipeline before saving terms to your system.
• Detect LLM hallucinations in extracted clauses (e.g., dates, parties, payment terms) by confirming presence and correctness in the contract text.
• Route only risky fields to legal or ops review while letting the rest pass automatically.

Technical details

• Judgment API integration via the community node n8n-nodes-judgment.
• Setup steps: install the node (Settings → Community Nodes), create a Judgment API credential, and configure provider Base URL if required.
• Workflow nodes include: if, set, no op, merge, sticky note, and n8n-nodes-judgment (Judgment evaluation).
• Customize by replacing the sample source_text and extracted fields with your own extraction output, then updating Judgment questions to include two checks per field (not-in-source and wrong-value).