LegalHalluLens: Typed Hallucination Auditing and Calibrated Multi-Agent Debate for Trustworthy Legal AI

LegalHalluLens is a new auditing framework designed to address hallucination issues in legal AI systems, which report an average hallucination rate of ~52%. It includes typed hallucination profiles for four claim categories, a Risk Direction Index (RDI) for bias reduction, and a calibrated debate pipeline that reduces fabricated detections by 45% while using a smaller model backbone of 4 billion parameters. This framework enhances accountability and agent design in legal AI deployments by revealing failure modes that traditional metrics obscure, thus enabling more trustworthy AI integration in legal workflows.