AnttiHero/lavern

What it solves

Lavern is a multi-agent system designed for legal document analysis. It moves away from the "AI as a junior associate" model, instead using a coordinated team of specialized agents to read documents, identify findings, and verify them through a rigorous debate and verification process to ensure high-quality, evidence-based legal work product.

How it works

The system employs a team of 67 specialized agent prompts (59 specialists and 8 orchestrators) running on frontier LLMs like Claude, Mistral, or local models via Ollama. It operates through several key mechanisms:

• Debate Protocol: Agents must provide specific citations from the document for any finding to be accepted. They can challenge each other's findings based on evidence.
• Three-Layer Verification: Findings pass through an evaluator gate, an adversarial debate (red team vs. blue team), and a 10-pass verification pipeline covering aspects like accuracy, risk, and legal design.
• Human Gates: The system pauses for human approval or overrides before critical decisions are finalized.
• Precedent Board: A persistent memory system that tracks recurring findings across different documents, promoting them from "tentative" to "confirmed" as they recur.

Who it’s for

Legal professionals and developers interested in multi-agent orchestration for high-stakes document review and analysis where auditability and verification are more critical than speed.

Highlights

• Extensive Agent Library: Includes 67 specialized prompts and 21 MCP tools for grounding and memory.
• Flexible Inference: Supports Anthropic, Mistral (for EU data residency), and local execution via Ollama.
• Autonomous Mode (Clawern): A daemon that monitors folders for new documents and pushes findings to Telegram or email.
• Auditability: Produces a full audit bundle including structured findings, debate resolutions, and cost logs.
• Comprehensive Testing: Includes over 1,600 tests and 5 seeded legal datasets (CUAD, MAUD, etc.).