EviBound: New Framework Eliminates False Claims in Autonomous AI Research

Researchers at Cornell University, led by Ruiying Chen, have introduced EviBound, an innovative governance framework designed to eliminate false claims in autonomous AI research agents. The framework addresses the 'integrity gap' prevalent in current autonomous systems, where confident claims are often made without verifiable evidence. EviBound employs a 'dual-gate' architecture that mandates machine-checkable evidence for all claims. The pre-execution Approval Gate validates acceptance criteria schemas before code runs, proactively catching structural violations. The post-execution Verification Gate validates artifacts via MLflow API queries, ensuring that claims are backed by a queryable run_id, required artifacts, and a 'FINISHED' status. Bounded, confidence-gated retries help recover from transient failures without unbounded loops. The framework was evaluated on eight benchmark tasks spanning infrastructure validation, ML capabilities, and governance stress tests. Results showed that Baseline A (Prompt-Level Only) yielded 100% hallucination, while Baseline B (Verification-Only) reduced it to 25%. EviBound, with its dual-gate system, achieved 0% hallucination, verifying 7/8 tasks and correctly blocking one at the approval gate with only approximately 8.3% execution overhead. This research emphasizes that research integrity is an architectural property achieved through governance gates rather than emerging from model scale. The reproducibility package includes execution trajectories, MLflow run_ids for all verified tasks, and a four-step verification protocol. The team suggests future work should focus on scaling to complex multi-step research, developing domain-specific evidence schemas, adaptive verification thresholds, cross-cycle learning from verification patterns, and planning and reflection system integration.