A new framework has been developed to diagnose and repair failures in reasoning language models, focusing on instruction hierarchy compliance. The proposed self-monitoring mechanisms significantly improve compliance rates across various models, indicating potential advancements in AI reliability and performance.