Stage-1 cold-start checkpoint from Ch1nyzzz/e2-learning: full-parameter online mistake-driven experience learning (next-observation SFT gated by a semantic judge) on ALFWorld AlfredTWEnv, initialized from Qwen/Qwen3-8B. Source run artifact: outputs/alfworld_qwen3_8b_active_10000_hf/checkpoints/env_step_006220.

Intended use: initialization ("cold start") for Stage-2 policy GRPO training with verl-agent. Load as a regular Hugging Face checkpoint (AutoModelForCausalLM.from_pretrained). controller_state.json is experiment-controller provenance metadata and can be ignored by inference/training stacks.