New LEDGER system audits LLM agents with trace graphs
LLM agents' work now auditable via layered claim-to-evidence graphs
A team from Lawrence Livermore National Laboratory (LLNL) has developed LEDGER (Layered Evidence and Decision Graphs for Execution Review), a tracing and review system designed to audit complex workflows performed by LLM agents.
As LLM agents increasingly handle long-horizon tasks involving tool use, code execution, and artifact generation, verifying the correctness and trustworthiness of their outputs has become critical. LEDGER addresses this by constructing layered trace graphs that map agent actions to specific claims, supporting evidence, and validation steps. The system organizes execution records into Evidence Nodes and Workflow Nodes, with semantic edges linking claims to supporting artifacts and checks. In case studies involving data analysis and coding, LEDGER successfully exposed workflow decisions, artifact lineage, repair steps, validation coverage, and claim-support paths—key elements for evidence-centered audits.
- LEDGER was developed by researchers at Lawrence Livermore National Laboratory (Daejong Kim, Haichao Miao, Shusen Liu)
- The system builds layered trace graphs to map agent claims to supporting evidence and validation steps
- Case studies show LEDGER exposes workflow decisions, artifact lineage, and repair steps for auditing
Why It Matters
Provides auditors with transparent, traceable evidence for verifying LLM agent outputs in critical workflows