B首更新前判分缺失审计

{
  "schema": 1,
  "kind": "credit_first_update_preflight_log_audit",
  "observed_at": "2026-10-01T15:35:00+08:00",
  "scope": "Counts of retained launch-log events before each arm first policy_entropy record; not a reconstructed training batch",
  "counts": {
    "O": {
      "first_metric_line": 4092,
      "format_errors_before_first_metric": 57,
      "p2p_regressions": 89,
      "no_reward_file": 1,
      "unjudged": 0
    },
    "B": {
      "first_metric_line": 3955,
      "format_errors_before_first_metric": 48,
      "p2p_regressions": 100,
      "no_reward_file": 1,
      "unjudged": 203
    },
    "M": {
      "first_metric_line": 4196,
      "format_errors_before_first_metric": 59,
      "p2p_regressions": 110,
      "no_reward_file": 0,
      "unjudged": 8
    }
  },
  "logs": {
    "O": {
      "path": "rl/runs/skyrl-qwen35-9b/rl9b-cfcredit-clean-O-b16n8-lr1e-6-fe05-p2p025-20260920/logs/launch_20260920_212653.log",
      "sha256": "d6678ae1a8fadddff010c696692a4a867328db265447ae46c7d6b7421fb7f26c"
    },
    "B": {
      "path": "rl/runs/skyrl-qwen35-9b/rl9b-cfcredit-clean-B-b16n8-lr1e-6-fe05-p2p025-20260920/logs/launch_20260922_001146.log",
      "sha256": "2b51f98b0da35285e73b6e3e1b8d1f47f70719b57346140dc3e93277c0c8d6cc"
    },
    "M": {
      "path": "rl/runs/skyrl-qwen35-9b/rl9b-cfcredit-clean-M-retry1-b16n8-lr1e-6-fe05-p2p025-20260924/logs/launch_20260924_100931.log",
      "sha256": "50de812d3cd9b3c5dc2c06f8d0695873c05190f9ad9f7d8cdf77300c393d4a26"
    }
  },
  "finding": "B had 203 no-judgement events before its first entropy record, and all 203 are Docker address-pool exhaustion errors. O had none; M had 8 BuildKit context-deadline errors. This makes B first-update entropy/reward a contaminated comparison for credit allocation, not evidence about broadcast credit.",
  "next_step": "Instrument a controlled first-update capture or rerun B under the same fixed profile; do not infer low entropy from online reward alone.",
  "failure_classification": {
    "O": {
      "unjudged_before_first_entropy": 0,
      "dominant_cause": null
    },
    "B": {
      "unjudged_before_first_entropy": 203,
      "exception_type": "RuntimeError",
      "cause": "Docker all predefined address pools have been fully subnetted"
    },
    "M": {
      "unjudged_before_first_entropy": 8,
      "exception_type": "RuntimeError",
      "cause": "BuildKit failed to solve: context deadline exceeded"
    }
  },
  "decision": "Treat original B80 as contaminated for the primary M-vs-B claim; rerun B under a clean network pool with gc/cleanup guard before using it as the control."
}