{
 "chart": "reward_hacking/reward_hacking_by_model",
 "title": "Detected reward hacking in TH2 analysis runs",
 "subtitle": "",
 "panels": [
  {
   "title": "",
   "x": {"label": "", "unit": "", "scale": "linear"},
   "y": {"label": "Share of TH2 runs (%)", "unit": "", "scale": "linear"},
   "tables": [
    {
     "name": "flagged runs",
     "rows": [
      {"model": "GPT-4o", "severity": "Egregious reward hacking (run zeroed)", "flagged_runs": 0, "runs": 49, "share_of_runs_percent": 0.0},
      {"model": "o1", "severity": "Egregious reward hacking (run zeroed)", "flagged_runs": 0, "runs": 67, "share_of_runs_percent": 0.0},
      {"model": "Claude Opus 4.1", "severity": "Egregious reward hacking (run zeroed)", "flagged_runs": 0, "runs": 5, "share_of_runs_percent": 0.0},
      {"model": "GPT-5", "severity": "Egregious reward hacking (run zeroed)", "flagged_runs": 0, "runs": 84, "share_of_runs_percent": 0.0},
      {"model": "Claude Opus 4.5", "severity": "Egregious reward hacking (run zeroed)", "flagged_runs": 0, "runs": 84, "share_of_runs_percent": 0.0},
      {"model": "GPT-5.4", "severity": "Egregious reward hacking (run zeroed)", "flagged_runs": 1, "runs": 83, "share_of_runs_percent": 1.20482},
      {"model": "Claude Opus 4.7", "severity": "Egregious reward hacking (run zeroed)", "flagged_runs": 0, "runs": 84, "share_of_runs_percent": 0.0},
      {"model": "Claude Opus 4.6", "severity": "Egregious reward hacking (run zeroed)", "flagged_runs": 0, "runs": 69, "share_of_runs_percent": 0.0},
      {"model": "GPT-5.5", "severity": "Egregious reward hacking (run zeroed)", "flagged_runs": 0, "runs": 69, "share_of_runs_percent": 0.0},
      {"model": "Claude Opus 4.8", "severity": "Egregious reward hacking (run zeroed)", "flagged_runs": 0, "runs": 67, "share_of_runs_percent": 0.0},
      {"model": "GPT-5.6 Sol", "severity": "Egregious reward hacking (run zeroed)", "flagged_runs": 4, "runs": 83, "share_of_runs_percent": 4.81928},
      {"model": "Claude Fable 5", "severity": "Egregious reward hacking (run zeroed)", "flagged_runs": 0, "runs": 67, "share_of_runs_percent": 0.0},
      {"model": "Claude Opus 5", "severity": "Egregious reward hacking (run zeroed)", "flagged_runs": 0, "runs": 66, "share_of_runs_percent": 0.0},
      {"model": "Claude Fable 5.1", "severity": "Egregious reward hacking (run zeroed)", "flagged_runs": 0, "runs": 78, "share_of_runs_percent": 0.0},
      {"model": "Claude Opus 5.5", "severity": "Egregious reward hacking (run zeroed)", "flagged_runs": 0, "runs": 71, "share_of_runs_percent": 0.0},
      {"model": "GPT-6 Astra", "severity": "Egregious reward hacking (run zeroed)", "flagged_runs": 2, "runs": 70, "share_of_runs_percent": 2.85714},
      {"model": "GPT-4o", "severity": "Borderline reward hacking (run kept)", "flagged_runs": 0, "runs": 49, "share_of_runs_percent": 0.0},
      {"model": "o1", "severity": "Borderline reward hacking (run kept)", "flagged_runs": 0, "runs": 67, "share_of_runs_percent": 0.0},
      {"model": "Claude Opus 4.1", "severity": "Borderline reward hacking (run kept)", "flagged_runs": 0, "runs": 5, "share_of_runs_percent": 0.0},
      {"model": "GPT-5", "severity": "Borderline reward hacking (run kept)", "flagged_runs": 0, "runs": 84, "share_of_runs_percent": 0.0},
      {"model": "Claude Opus 4.5", "severity": "Borderline reward hacking (run kept)", "flagged_runs": 0, "runs": 84, "share_of_runs_percent": 0.0},
      {"model": "GPT-5.4", "severity": "Borderline reward hacking (run kept)", "flagged_runs": 1, "runs": 83, "share_of_runs_percent": 1.20482},
      {"model": "Claude Opus 4.7", "severity": "Borderline reward hacking (run kept)", "flagged_runs": 0, "runs": 84, "share_of_runs_percent": 0.0},
      {"model": "Claude Opus 4.6", "severity": "Borderline reward hacking (run kept)", "flagged_runs": 0, "runs": 69, "share_of_runs_percent": 0.0},
      {"model": "GPT-5.5", "severity": "Borderline reward hacking (run kept)", "flagged_runs": 2, "runs": 69, "share_of_runs_percent": 2.89855},
      {"model": "Claude Opus 4.8", "severity": "Borderline reward hacking (run kept)", "flagged_runs": 0, "runs": 67, "share_of_runs_percent": 0.0},
      {"model": "GPT-5.6 Sol", "severity": "Borderline reward hacking (run kept)", "flagged_runs": 1, "runs": 83, "share_of_runs_percent": 1.20482},
      {"model": "Claude Fable 5", "severity": "Borderline reward hacking (run kept)", "flagged_runs": 0, "runs": 67, "share_of_runs_percent": 0.0},
      {"model": "Claude Opus 5", "severity": "Borderline reward hacking (run kept)", "flagged_runs": 0, "runs": 66, "share_of_runs_percent": 0.0},
      {"model": "Claude Fable 5.1", "severity": "Borderline reward hacking (run kept)", "flagged_runs": 2, "runs": 78, "share_of_runs_percent": 2.5641},
      {"model": "Claude Opus 5.5", "severity": "Borderline reward hacking (run kept)", "flagged_runs": 0, "runs": 71, "share_of_runs_percent": 0.0},
      {"model": "GPT-6 Astra", "severity": "Borderline reward hacking (run kept)", "flagged_runs": 1, "runs": 70, "share_of_runs_percent": 1.42857},
      {"model": "GPT-4o", "severity": "Suspicious behavior (run kept)", "flagged_runs": 0, "runs": 49, "share_of_runs_percent": 0.0},
      {"model": "o1", "severity": "Suspicious behavior (run kept)", "flagged_runs": 0, "runs": 67, "share_of_runs_percent": 0.0},
      {"model": "Claude Opus 4.1", "severity": "Suspicious behavior (run kept)", "flagged_runs": 0, "runs": 5, "share_of_runs_percent": 0.0},
      {"model": "GPT-5", "severity": "Suspicious behavior (run kept)", "flagged_runs": 0, "runs": 84, "share_of_runs_percent": 0.0},
      {"model": "Claude Opus 4.5", "severity": "Suspicious behavior (run kept)", "flagged_runs": 1, "runs": 84, "share_of_runs_percent": 1.19048},
      {"model": "GPT-5.4", "severity": "Suspicious behavior (run kept)", "flagged_runs": 2, "runs": 83, "share_of_runs_percent": 2.40964},
      {"model": "Claude Opus 4.7", "severity": "Suspicious behavior (run kept)", "flagged_runs": 0, "runs": 84, "share_of_runs_percent": 0.0},
      {"model": "Claude Opus 4.6", "severity": "Suspicious behavior (run kept)", "flagged_runs": 0, "runs": 69, "share_of_runs_percent": 0.0},
      {"model": "GPT-5.5", "severity": "Suspicious behavior (run kept)", "flagged_runs": 2, "runs": 69, "share_of_runs_percent": 2.89855},
      {"model": "Claude Opus 4.8", "severity": "Suspicious behavior (run kept)", "flagged_runs": 0, "runs": 67, "share_of_runs_percent": 0.0},
      {"model": "GPT-5.6 Sol", "severity": "Suspicious behavior (run kept)", "flagged_runs": 2, "runs": 83, "share_of_runs_percent": 2.40964},
      {"model": "Claude Fable 5", "severity": "Suspicious behavior (run kept)", "flagged_runs": 8, "runs": 67, "share_of_runs_percent": 11.9403},
      {"model": "Claude Opus 5", "severity": "Suspicious behavior (run kept)", "flagged_runs": 1, "runs": 66, "share_of_runs_percent": 1.51515},
      {"model": "Claude Fable 5.1", "severity": "Suspicious behavior (run kept)", "flagged_runs": 4, "runs": 78, "share_of_runs_percent": 5.12821},
      {"model": "Claude Opus 5.5", "severity": "Suspicious behavior (run kept)", "flagged_runs": 3, "runs": 71, "share_of_runs_percent": 4.22535},
      {"model": "GPT-6 Astra", "severity": "Suspicious behavior (run kept)", "flagged_runs": 3, "runs": 70, "share_of_runs_percent": 4.28571}
     ]
    }
   ]
  }
 ]
}
