A new framework called EviBound has been developed to address the problem of false claims in autonomous AI research. Created by Ruiying Chen at Cornell University, EviBound enforces dual governance gates that require machine-checkable evidence, drastically reducing and even eliminating instances of AI hallucination. The core of EviBound's innovation is its dual-gate architecture:

Approval Gate: Validates acceptance criteria schemas before code execution, proactively catching structural violations.
Verification Gate: Validates artifacts after execution via MLflow API queries, ensuring that claimed results exist and match the acceptance criteria.

Evaluations of the framework across benchmark tasks demonstrated a significant reduction in hallucination rates. Baseline systems relying solely on prompt-level techniques exhibited 100% hallucination, while systems using only verification reduced it to 25%. EviBound, employing both gates, achieved 0% hallucination with minimal execution overhead. This framework includes execution trajectories, MLflow run IDs for all verified tasks, and a four-step verification protocol. Research integrity is achieved through governance gates rather than model scale. Key findings emphasize that architectural enforcement, not model size, is critical for trustworthy autonomous research. The system's design provides a reusable benchmark and architectural template for future developments in AI research.