What is "multi-turn rollout locking" in the blog?

[Weili-0234]

The blog says that

To continuously unleash the power of reinforcement learning, we built a scalable asynchronous RL framework that supports Qwen3.5 models of all sizes, spanning text, multimodal, and multi-turn settings. By adopting a fully disaggregated training-inference architecture, the framework achieves significantly improved hardware utilization, dynamic load balancing, and fine-grained fault recovery. It further optimizes throughput and enhances train–infer consistency via techniques such as FP8 end-to-end training, rollout router replay, speculative decoding, and multi-turn rollout locking.

It looks like this is a relatively new concept for the RL infra community. Could you further explain what this means? Or is there going to be a technical report that covers this?