EviBound: AI Governance Framework Eliminates False Claims in Autonomous Research

Cornell University researcher Ruiying Chen has introduced EviBound, a governance framework designed to eliminate false claims generated by LLM-based autonomous research agents. The framework addresses the challenge of 'hallucinations'—instances where AI systems confidently report results unsupported by verifiable evidence. EviBound employs a dual-gate architecture requiring machine-checkable evidence for all claims. The Approval Gate validates acceptance criteria schemas before code execution, proactively identifying structural violations. The Verification Gate validates artifacts post-execution using MLflow API queries, ensuring claimed results are backed by queryable run_ids, required artifacts, and a 'FINISHED' status. The framework was evaluated across eight benchmark tasks spanning infrastructure validation, ML capabilities, and governance stress tests. EviBound achieved 0% hallucination compared to a 100% hallucination rate in systems relying solely on prompt-level controls and a 25% rate when using verification alone. The system's design ensures research integrity as an architectural property rather than an emergent result of model scale. Key features of EviBound include bounded, confidence-gated retries to recover from transient failures, MLflow integration for artifact validation, and a claims ledger that provides full provenance for verified tasks. This system provides a benchmark and architectural template for developing trustworthy autonomous research systems. The research indicates that architectural enforcement, not merely model scale, is critical for maintaining integrity in autonomous research. The reproducibility package, including execution trajectories and MLflow run_ids, will be made available upon publication under CC BY 4.0 licensing.