Lenz

Lenz is an audit-grade fact-checking API built for AI product teams that need to catch hallucinations and factual errors before they reach users. It runs every claim through a structured five-stage pipeline — framing, research, debate, panel review, and conclusion — using eight models from multiple providers, grounded in retrieved sources with a full citation trail. The product is accessible via REST API, Python SDK, TypeScript SDK, and integrations with tools like Zapier, n8n, and Claude's MCP server.

What It Is

Lenz is a multi-vendor LLM fact-checking service that sits between an AI product's output and its end users. Rather than asking a single model whether a claim is true, Lenz frames the claim as a falsifiable statement, retrieves independent sources, runs opposing-side debate between models, convenes a three-reviewer panel, and returns a structured verdict with a score, sourced citations, and a full reasoning trace. The output is machine-readable so downstream workflows can branch automatically — block, publish, or escalate.

The Five-Stage Pipeline

The pipeline is the core architectural differentiator Lenz describes on its about and how-it-works pages:

• Framing — The claim is rewritten as a precise, falsifiable statement before research begins.
• Research — Multiple queries run across authoritative sources to retrieve evidence.
• Debate — Two models argue opposing sides from the retrieved evidence; disagreement is preserved, not averaged away.
• Panel Review — Three independent reviewers audit source reliability, logical fallacies, and claim precision on separate axes.
• Conclusion — A final verdict, score, and citation trail are returned in structured output.

API Surface and Integrations

Lenz exposes four API endpoints that form what the product page calls a "depth ladder":

• /extract — Pull verifiable claims from any text
• /assess — Fast multi-model inline guardrail for synchronous UX
• /verify — Full eight-model pipeline from framing to sourced verdict
• /ask — Ask a factual question and receive a sourced answer

Integrations include Zapier, n8n, Claude (via MCP server), a REST API, a Python SDK, and a TypeScript SDK. The product page notes it is cited by ChatGPT, Perplexity, and Gemini, and runs inside Claude, Cursor, n8n, and Zapier.

Target Use Cases

Lenz identifies six primary integration patterns on its product page:

• AI content before publishing — Newsletters, blog posts, and product updates verified before going live
• Data enrichment and CRM — Catching invented funding rounds, headcounts, or tech stacks before they enter a CRM
• Regulated claims — Health, finance, and supplement copy where the audit trail is the deliverable
• Agent guardrail — Wiring /assess as a pre-action check before an agent acts on a factual claim
• Customer support AI — Catching wrong answers before they reach customers
• RAG response gating — Groundedness checks verifying retrieval faithfulness against the open web

Why It Matters: The Model Disagreement Problem

Lenz publishes research on LLM disagreement that underpins its multi-model approach. According to the company's own study of five frontier models each ruling on 1,000 real claims, at least one model broke from the panel majority — or no majority formed at all — on 63% of claims, and 23% had a substantial split of two or more verdict options apart. The company frames this as evidence that a single confident model answer can hide significant uncertainty, and that a panel approach grounded in retrieved sources is necessary to produce a defensible verdict.