CrossAudit Ensures Trustworthy AI Scientific Research

This paper introduces CrossAudit, a novel protocol designed to ensure the integrity and auditability of autonomous scientific research pipelines powered by AI agents. The core premise is to prevent AI systems from self-validating their own work, addressing concerns that models from the same family or vendor might share blind spots. CrossAudit establishes three key commitments: every piece of work is audited by an agent from a different vendor against a human-defined, version-controlled rulebook; all supervision history, including reports, verdicts, and disputes, is recorded as Git commits for transparency and traceability; and automated, scripted checks run before any model intervention, with advisory judgments never gating the pipeline unless a rule is explicitly cited. The protocol is defined by eight invariants, and a reference implementation using GitHub Actions and Python is described, along with a live deployment in a computational chemistry pipeline. A seeded-defect trial demonstrated that different vendors interpret the same rulebook differently, underscoring the value of cross-vendor auditing. The paper itself underwent a cross-vendor audit, with its findings incorporated, highlighting the practical application and commitment to the protocol's principles. This system aims to provide a robust, transparent, and verifiable method for overseeing AI-driven scientific discovery.