# Lenz

> An independent, multi-model fact-checking API that verifies AI-generated claims against real sources using an eight-model, five-stage adversarial pipeline with full citation trails.

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.

## Features
- Eight-model, five-stage fact-checking pipeline
- Opposing-side adversarial debate between models
- Three independent panel reviewers
- Full citation trail with sourced verdicts
- Machine-readable structured output (block, publish, or escalate)
- Claim framing to remove ambiguity before research
- /extract endpoint to pull verifiable claims from text
- /assess endpoint for fast inline guardrail
- /verify endpoint for full adversarial pipeline
- /ask endpoint for sourced factual Q&A
- Python SDK
- TypeScript SDK
- REST API with OpenAPI 3.1
- Zapier integration
- n8n integration
- Claude MCP server integration
- Pre-release CI hallucination regression testing
- Runtime gate for outbound AI text
- Incident triage with evidence and citation trail
- RAG response groundedness checking

## Integrations
Zapier, n8n, Claude (MCP server), Cursor, REST API, Python SDK, TypeScript SDK

## Platforms
API, WEB

## Pricing
Freemium — Free tier available with paid upgrades

## Links
- Website: https://lenz.io/
- Documentation: https://lenz.io/developers
- EveryDev.ai: https://www.everydev.ai/tools/lenz