Ian Barber describes several concrete sources of divergence. Observed facts: The post says QwenGyre rollouts for Alibaba’s 122B model averaged around two hours, while many runs for its 2.4T flagship exceeded four hours. It also says QwenGyre caches original tool-call payloads because the harness reformats them between turns. DeepSeek, meanwhile, moved its harness into the sandbox in version 4.1 and froze state including DRAM and the writable filesystem layer when preempting work.
Mechanism: Long sessions accumulate opportunities for mismatch across model output, inference engine, harness, sandbox, tool-call parsing, branch serialization, and training kernels. A session that includes subagents and parallel calls is a tree, not a simple token stream; serialization and weighting choices affect what the trainer learns. Why now: Agentic post-training increasingly uses long, tool-rich runs, making these operational details consequential rather than incidental.
First implication: Teams building agent-training pipelines should treat trajectory fidelity—including state, tool payloads, branch structure, and compute parity—as part of the training system’s correctness budget. The source describes multiple remedies, including recomputation, ratio correction, and bitwise consistency, while noting performance costs; it does not establish one best approach. Evidence boundary: This is an engineering analysis with examples from cited systems, not a controlled benchmark comparing all remedies. First watch condition: Watch whether future training reports quantify how trajectory-preservation techniques affect both policy quality and throughput.