Logic-Graph Verification System

Short summary

A verification system for cheap language models that externalizes reasoning as directed graphs, samples outputs N times independently, and uses mechanical graph checks to identify conclusions with structural support. Rather than asking models to self-correct or verify themselves, this system tracks argument topology: conclusions backed by multiple independent reasoning chains are more trustworthy than single-path assertions. Intended for document-based Q&A, claim verification, and as an MCP tool for agent working memory.

• •Externalize LLM reasoning as graphs and verify structure mechanically instead of asking models to judge themselves
• •Sample independent model runs; conclusions supported by multiple disjoint argument paths are more reliable than single paths
• •Designed for cheap open-weight models via OpenRouter, with library and MCP deployment modes for agents

Generated with AI, which can make mistakes.