Researchers have identified a critical issue with on-policy distillation (OPD) in large language models, where training data becomes dominated by truncated trajectories due to abrupt length inflation. As student models learn from their own induced distribution and stronger teachers, they can experience a failure mode characterized by repetition saturation and truncation collapse. This phenomenon occurs when on-policy rollouts undergo sudden increases in length, causing the training data to become unrepresentative of the intended distribution. The consequences of this issue can be severe, leading to suboptimal performance and instability in the trained models. To address this problem, stabilization strategies are necessary to prevent length inflation and ensure that the training data remains diverse and representative. The discovery of this failure mode has significant implications for the development of reliable and efficient large language models, so what matters most to practitioners is the need to implement effective stabilization techniques to mitigate the risks associated with OPD1.