B受控重跑预检计划（未放行）

{
  "schema": 1,
  "kind": "B_clean_rerun_preflight_plan",
  "created_at": "2026-10-01T15:45:00+08:00",
  "status": "not_launch_ready",
  "reason": "gpu23 external VLLM occupies all 8 GPUs; first-update instrumentation must be added before launch",
  "run_name": "rl9b-cfcredit-clean-B-rerun1-b16n8-lr1e-6-fe05-p2p025-20261001",
  "launcher": "rl/skyrl/scripts/qwen35-9b/run_rl9b_cfcredit_clean_20260920.sh",
  "env": {
    "ARM": "B",
    "RUN_NAME": "rl9b-cfcredit-clean-B-rerun1-b16n8-lr1e-6-fe05-p2p025-20261001",
    "TRAIN_DATA_DIR": "/data1/home/zly/swe/rl/data/sbvpool-20260920-clean-train2400",
    "HARBOR_TRIAL_CONFIG_FILE": "mini_swe_tito_clean",
    "RESUME_MODE": "none",
    "MAX_TRAINING_STEPS": "80"
  },
  "must_pass_before_launch": [
    "gpu08 GPU0-5 free; preserve GPU6/7 and small service",
    "gpu23 GPU0-7 free and externally released",
    "docker network count at baseline 3 before launch",
    "exact cleanup of prior B trial leftovers",
    "start gc_orphan_networks.sh guard",
    "capture first consumed batch: response_mask token count, entropy numerator/denominator, reward/status counts, no-judgement causes",
    "same source/profile hashes as original O/M clean runs"
  ],
  "success_criteria": [
    "first update has batch-level receipt",
    "no address-pool exhaustion before first update",
    "B first-step entropy is interpretable under identical response-mask accounting",
    "80 updates, checkpoint/export step80, exact cleanup"
  ],
  "do_not_use_old_B80_for_primary_claim": true
}