{
 "arm": "base",
 "tag": "base",
 "think": true,
 "K": 12,
 "budget": 2048,
 "answer_max": 512,
 "temperature": 0.8,
 "top_p": 0.95,
 "adapter": null,
 "n_tasks": 60,
 "rows": [
  {
   "task_id": 1,
   "family": "list",
   "depth": 3,
   "prompt_tokens": 275,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2075,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2083,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 591,
     "n_think": 560,
     "forced": false
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 968,
     "n_think": 931,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2086,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    }
   ]
  },
  {
   "task_id": 2,
   "family": "list",
   "depth": 3,
   "prompt_tokens": 361,
   "greedy": {
    "parsed": false,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2562,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": true,
     "probe_robust": true,
     "mimicry": false,
     "full": true,
     "gen_tokens": 2077,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1111,
     "n_think": 1081,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 675,
     "n_think": 580,
     "forced": false
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 768,
     "n_think": 719,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": true,
     "probe_robust": true,
     "mimicry": false,
     "full": true,
     "gen_tokens": 2126,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1559,
     "n_think": 1448,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2127,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": true,
     "probe_robust": true,
     "mimicry": false,
     "full": true,
     "gen_tokens": 2104,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    }
   ]
  },
  {
   "task_id": 3,
   "family": "list",
   "depth": 3,
   "prompt_tokens": 322,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2095,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2129,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2157,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1395,
     "n_think": 1332,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 884,
     "n_think": 849,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2109,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2125,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1395,
     "n_think": 1360,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1743,
     "n_think": 1690,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1045,
     "n_think": 985,
     "forced": false
    }
   ]
  },
  {
   "task_id": 4,
   "family": "list",
   "depth": 3,
   "prompt_tokens": 437,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2077,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2076,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 619,
     "n_think": 593,
     "forced": false
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2074,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2393,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 746,
     "n_think": 720,
     "forced": false
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2129,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    }
   ]
  },
  {
   "task_id": 5,
   "family": "list",
   "depth": 3,
   "prompt_tokens": 386,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2116,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2092,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2093,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2110,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2088,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2088,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2101,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    }
   ]
  },
  {
   "task_id": 6,
   "family": "list",
   "depth": 3,
   "prompt_tokens": 340,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2562,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": true,
     "probe_robust": true,
     "mimicry": false,
     "full": true,
     "gen_tokens": 2102,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 849,
     "n_think": 744,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2093,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2114,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2147,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": true,
     "probe_robust": true,
     "mimicry": false,
     "full": true,
     "gen_tokens": 2078,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    }
   ]
  },
  {
   "task_id": 7,
   "family": "list",
   "depth": 3,
   "prompt_tokens": 352,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2382,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1443,
     "n_think": 1409,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1299,
     "n_think": 1047,
     "forced": false
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2130,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 848,
     "n_think": 788,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1290,
     "n_think": 1205,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 623,
     "n_think": 595,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1204,
     "n_think": 1163,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2112,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1627,
     "n_think": 1497,
     "forced": false
    }
   ]
  },
  {
   "task_id": 8,
   "family": "list",
   "depth": 3,
   "prompt_tokens": 293,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2102,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2138,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2123,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2137,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2125,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1105,
     "n_think": 1009,
     "forced": false
    }
   ]
  },
  {
   "task_id": 9,
   "family": "list",
   "depth": 3,
   "prompt_tokens": 372,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2173,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1357,
     "n_think": 1224,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1131,
     "n_think": 1007,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2092,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2167,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 688,
     "n_think": 633,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1170,
     "n_think": 1140,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1446,
     "n_think": 932,
     "forced": false
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1474,
     "n_think": 1424,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2122,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    }
   ]
  },
  {
   "task_id": 10,
   "family": "list",
   "depth": 3,
   "prompt_tokens": 425,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2562,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2085,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": true,
     "probe_robust": true,
     "mimicry": false,
     "full": true,
     "gen_tokens": 2127,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2086,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2143,
     "n_think": 2048,
     "forced": true
    }
   ]
  },
  {
   "task_id": 11,
   "family": "list",
   "depth": 3,
   "prompt_tokens": 385,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2102,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1047,
     "n_think": 800,
     "forced": false
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2102,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2087,
     "n_think": 2048,
     "forced": true
    }
   ]
  },
  {
   "task_id": 12,
   "family": "list",
   "depth": 3,
   "prompt_tokens": 403,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2118,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2117,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2086,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    }
   ]
  },
  {
   "task_id": 13,
   "family": "list",
   "depth": 3,
   "prompt_tokens": 397,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2084,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": true,
     "probe_robust": true,
     "mimicry": false,
     "full": true,
     "gen_tokens": 1551,
     "n_think": 665,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": true,
     "probe_robust": false,
     "mimicry": true,
     "full": false,
     "gen_tokens": 2317,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2121,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": true,
     "probe_robust": true,
     "mimicry": false,
     "full": true,
     "gen_tokens": 2321,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": true,
     "probe_robust": false,
     "mimicry": true,
     "full": false,
     "gen_tokens": 2180,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": true,
     "probe_robust": true,
     "mimicry": false,
     "full": true,
     "gen_tokens": 2256,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": true,
     "probe_robust": true,
     "mimicry": false,
     "full": true,
     "gen_tokens": 2083,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    }
   ]
  },
  {
   "task_id": 14,
   "family": "list",
   "depth": 3,
   "prompt_tokens": 370,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2079,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1050,
     "n_think": 974,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2072,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2089,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2150,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1128,
     "n_think": 1058,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2463,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2106,
     "n_think": 2048,
     "forced": true
    }
   ]
  },
  {
   "task_id": 15,
   "family": "list",
   "depth": 3,
   "prompt_tokens": 500,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2562,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2224,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2112,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1374,
     "n_think": 1326,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2088,
     "n_think": 2048,
     "forced": true
    }
   ]
  },
  {
   "task_id": 16,
   "family": "list",
   "depth": 3,
   "prompt_tokens": 397,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2562,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2119,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1255,
     "n_think": 991,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1747,
     "n_think": 1688,
     "forced": false
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1160,
     "n_think": 1130,
     "forced": false
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2426,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 543,
     "n_think": 518,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2215,
     "n_think": 2048,
     "forced": true
    }
   ]
  },
  {
   "task_id": 17,
   "family": "list",
   "depth": 3,
   "prompt_tokens": 321,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2562,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1959,
     "n_think": 1754,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1188,
     "n_think": 1069,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2129,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1432,
     "n_think": 1332,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2099,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1779,
     "n_think": 1677,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    }
   ]
  },
  {
   "task_id": 18,
   "family": "list",
   "depth": 3,
   "prompt_tokens": 423,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2087,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1885,
     "n_think": 1371,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2086,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1308,
     "n_think": 1234,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2109,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2167,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    }
   ]
  },
  {
   "task_id": 19,
   "family": "list",
   "depth": 3,
   "prompt_tokens": 412,
   "greedy": {
    "parsed": false,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2562,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1887,
     "n_think": 1839,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 883,
     "n_think": 647,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2099,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 608,
     "n_think": 564,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1075,
     "n_think": 927,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2343,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2102,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2078,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1084,
     "n_think": 1020,
     "forced": false
    }
   ]
  },
  {
   "task_id": 20,
   "family": "list",
   "depth": 3,
   "prompt_tokens": 328,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 858,
    "n_think": 824,
    "forced": false
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1137,
     "n_think": 957,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2160,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 587,
     "n_think": 525,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2151,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1008,
     "n_think": 823,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1116,
     "n_think": 1084,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2165,
     "n_think": 1996,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2127,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1792,
     "n_think": 1761,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 806,
     "n_think": 662,
     "forced": false
    }
   ]
  },
  {
   "task_id": 21,
   "family": "list",
   "depth": 3,
   "prompt_tokens": 444,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 923,
    "n_think": 585,
    "forced": false
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2123,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2359,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1223,
     "n_think": 1188,
     "forced": false
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2396,
     "n_think": 2048,
     "forced": true
    }
   ]
  },
  {
   "task_id": 22,
   "family": "list",
   "depth": 3,
   "prompt_tokens": 367,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2090,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2123,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2101,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2164,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2125,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2113,
     "n_think": 2048,
     "forced": true
    }
   ]
  },
  {
   "task_id": 23,
   "family": "list",
   "depth": 3,
   "prompt_tokens": 397,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2114,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1284,
     "n_think": 1212,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2124,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 666,
     "n_think": 591,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    }
   ]
  },
  {
   "task_id": 24,
   "family": "list",
   "depth": 3,
   "prompt_tokens": 376,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2100,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2084,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2350,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2084,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1172,
     "n_think": 1044,
     "forced": false
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2082,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2089,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2100,
     "n_think": 2048,
     "forced": true
    }
   ]
  },
  {
   "task_id": 25,
   "family": "list",
   "depth": 3,
   "prompt_tokens": 355,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2103,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2168,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2128,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2132,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2117,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    }
   ]
  },
  {
   "task_id": 26,
   "family": "list",
   "depth": 3,
   "prompt_tokens": 396,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2073,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1536,
     "n_think": 1509,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1681,
     "n_think": 1656,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1127,
     "n_think": 1102,
     "forced": false
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2337,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1285,
     "n_think": 1260,
     "forced": false
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 612,
     "n_think": 585,
     "forced": false
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    }
   ]
  },
  {
   "task_id": 27,
   "family": "list",
   "depth": 3,
   "prompt_tokens": 266,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2068,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    }
   ]
  },
  {
   "task_id": 28,
   "family": "list",
   "depth": 3,
   "prompt_tokens": 464,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2232,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2090,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2113,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2097,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2081,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2113,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1545,
     "n_think": 1474,
     "forced": false
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    }
   ]
  },
  {
   "task_id": 29,
   "family": "list",
   "depth": 3,
   "prompt_tokens": 379,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2562,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2201,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2231,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1264,
     "n_think": 1141,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2202,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2076,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 697,
     "n_think": 593,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2159,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2003,
     "n_think": 1754,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2232,
     "n_think": 2048,
     "forced": true
    }
   ]
  },
  {
   "task_id": 30,
   "family": "list",
   "depth": 3,
   "prompt_tokens": 331,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2113,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 707,
     "n_think": 612,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2335,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1030,
     "n_think": 947,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2247,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2071,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2172,
     "n_think": 2048,
     "forced": true
    }
   ]
  },
  {
   "task_id": 1,
   "family": "string",
   "depth": 3,
   "prompt_tokens": 204,
   "greedy": {
    "parsed": false,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2562,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1137,
     "n_think": 1024,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1565,
     "n_think": 1341,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 720,
     "n_think": 646,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2200,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1042,
     "n_think": 978,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1026,
     "n_think": 544,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2197,
     "n_think": 2048,
     "forced": true
    }
   ]
  },
  {
   "task_id": 2,
   "family": "string",
   "depth": 3,
   "prompt_tokens": 171,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2088,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2127,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2133,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2314,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 927,
     "n_think": 673,
     "forced": false
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1232,
     "n_think": 988,
     "forced": false
    }
   ]
  },
  {
   "task_id": 3,
   "family": "string",
   "depth": 3,
   "prompt_tokens": 183,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2068,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2106,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2089,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 858,
     "n_think": 838,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2160,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1200,
     "n_think": 1180,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2104,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1932,
     "n_think": 1747,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 757,
     "n_think": 737,
     "forced": false
    }
   ]
  },
  {
   "task_id": 4,
   "family": "string",
   "depth": 3,
   "prompt_tokens": 166,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 698,
    "n_think": 676,
    "forced": false
   },
   "samples": [
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2096,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2091,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2221,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 839,
     "n_think": 810,
     "forced": false
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    }
   ]
  },
  {
   "task_id": 5,
   "family": "string",
   "depth": 3,
   "prompt_tokens": 175,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2562,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2120,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 693,
     "n_think": 610,
     "forced": false
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1477,
     "n_think": 1392,
     "forced": false
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2090,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1323,
     "n_think": 1136,
     "forced": false
    }
   ]
  },
  {
   "task_id": 6,
   "family": "string",
   "depth": 3,
   "prompt_tokens": 177,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2096,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2068,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1165,
     "n_think": 1038,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2142,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2077,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1354,
     "n_think": 1334,
     "forced": false
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 998,
     "n_think": 961,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2077,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1242,
     "n_think": 1213,
     "forced": false
    }
   ]
  },
  {
   "task_id": 7,
   "family": "string",
   "depth": 3,
   "prompt_tokens": 185,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2281,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1654,
     "n_think": 1608,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2127,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1262,
     "n_think": 1062,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2249,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2092,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2108,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1245,
     "n_think": 1214,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2266,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2094,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 554,
     "n_think": 523,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1010,
     "n_think": 914,
     "forced": false
    }
   ]
  },
  {
   "task_id": 8,
   "family": "string",
   "depth": 3,
   "prompt_tokens": 178,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2460,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1022,
     "n_think": 768,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 818,
     "n_think": 731,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1843,
     "n_think": 1725,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2322,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 888,
     "n_think": 624,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2223,
     "n_think": 2048,
     "forced": true
    }
   ]
  },
  {
   "task_id": 9,
   "family": "string",
   "depth": 3,
   "prompt_tokens": 184,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2137,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 940,
     "n_think": 829,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 998,
     "n_think": 826,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2270,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1570,
     "n_think": 1493,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2135,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2223,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1232,
     "n_think": 1085,
     "forced": false
    }
   ]
  },
  {
   "task_id": 10,
   "family": "string",
   "depth": 3,
   "prompt_tokens": 187,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2116,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2190,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 653,
     "n_think": 548,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 623,
     "n_think": 525,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 853,
     "n_think": 736,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1105,
     "n_think": 725,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2148,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2077,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2144,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 712,
     "n_think": 669,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    }
   ]
  },
  {
   "task_id": 11,
   "family": "string",
   "depth": 3,
   "prompt_tokens": 163,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 790,
    "n_think": 674,
    "forced": false
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1230,
     "n_think": 1201,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 740,
     "n_think": 690,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 965,
     "n_think": 936,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 670,
     "n_think": 615,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2082,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": true,
     "probe_robust": false,
     "mimicry": true,
     "full": false,
     "gen_tokens": 2084,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2077,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": true,
     "probe_robust": false,
     "mimicry": true,
     "full": false,
     "gen_tokens": 2095,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1550,
     "n_think": 1512,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2082,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1390,
     "n_think": 1315,
     "forced": false
    }
   ]
  },
  {
   "task_id": 12,
   "family": "string",
   "depth": 3,
   "prompt_tokens": 218,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2202,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1791,
     "n_think": 1277,
     "forced": false
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2104,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    }
   ]
  },
  {
   "task_id": 13,
   "family": "string",
   "depth": 3,
   "prompt_tokens": 190,
   "greedy": {
    "parsed": false,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2562,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2073,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 992,
     "n_think": 907,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2073,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2073,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 621,
     "n_think": 560,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2073,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2374,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2230,
     "n_think": 2048,
     "forced": true
    }
   ]
  },
  {
   "task_id": 14,
   "family": "string",
   "depth": 3,
   "prompt_tokens": 176,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2092,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1249,
     "n_think": 1227,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2135,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1047,
     "n_think": 1012,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2077,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2101,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1507,
     "n_think": 1365,
     "forced": false
    }
   ]
  },
  {
   "task_id": 15,
   "family": "string",
   "depth": 3,
   "prompt_tokens": 170,
   "greedy": {
    "parsed": true,
    "legacy": true,
    "probe_robust": true,
    "mimicry": false,
    "full": true,
    "gen_tokens": 910,
    "n_think": 880,
    "forced": false
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": true,
     "probe_robust": true,
     "mimicry": false,
     "full": false,
     "gen_tokens": 944,
     "n_think": 900,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": true,
     "probe_robust": true,
     "mimicry": false,
     "full": true,
     "gen_tokens": 1046,
     "n_think": 1016,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": true,
     "probe_robust": true,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2068,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1610,
     "n_think": 1517,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": true,
     "probe_robust": true,
     "mimicry": false,
     "full": true,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2119,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2085,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": true,
     "probe_robust": true,
     "mimicry": false,
     "full": false,
     "gen_tokens": 545,
     "n_think": 525,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": true,
     "probe_robust": true,
     "mimicry": false,
     "full": true,
     "gen_tokens": 2078,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": true,
     "probe_robust": true,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2094,
     "n_think": 2048,
     "forced": true
    }
   ]
  },
  {
   "task_id": 16,
   "family": "string",
   "depth": 3,
   "prompt_tokens": 173,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2070,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2025,
     "n_think": 1981,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2141,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1400,
     "n_think": 1365,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1402,
     "n_think": 1308,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1980,
     "n_think": 1923,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2105,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1970,
     "n_think": 1413,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2202,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1018,
     "n_think": 984,
     "forced": false
    }
   ]
  },
  {
   "task_id": 17,
   "family": "string",
   "depth": 3,
   "prompt_tokens": 174,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2113,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1066,
     "n_think": 1007,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1310,
     "n_think": 1129,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2113,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1143,
     "n_think": 1012,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2198,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2343,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 855,
     "n_think": 832,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2299,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    }
   ]
  },
  {
   "task_id": 18,
   "family": "string",
   "depth": 3,
   "prompt_tokens": 179,
   "greedy": {
    "parsed": false,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2562,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1739,
     "n_think": 1504,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1478,
     "n_think": 1290,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1554,
     "n_think": 753,
     "forced": false
    }
   ]
  },
  {
   "task_id": 19,
   "family": "string",
   "depth": 3,
   "prompt_tokens": 177,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 1391,
    "n_think": 1337,
    "forced": false
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1014,
     "n_think": 982,
     "forced": false
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2088,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2088,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2078,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2071,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2345,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 766,
     "n_think": 679,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": true,
     "probe_robust": false,
     "mimicry": true,
     "full": true,
     "gen_tokens": 2134,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2122,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": true,
     "probe_robust": false,
     "mimicry": true,
     "full": true,
     "gen_tokens": 2114,
     "n_think": 2048,
     "forced": true
    }
   ]
  },
  {
   "task_id": 20,
   "family": "string",
   "depth": 3,
   "prompt_tokens": 241,
   "greedy": {
    "parsed": true,
    "legacy": true,
    "probe_robust": false,
    "mimicry": true,
    "full": false,
    "gen_tokens": 2203,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2297,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": true,
     "probe_robust": false,
     "mimicry": true,
     "full": false,
     "gen_tokens": 1793,
     "n_think": 1651,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": true,
     "probe_robust": false,
     "mimicry": true,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": true,
     "probe_robust": true,
     "mimicry": false,
     "full": true,
     "gen_tokens": 2363,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": true,
     "probe_robust": false,
     "mimicry": true,
     "full": false,
     "gen_tokens": 2211,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1198,
     "n_think": 950,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1182,
     "n_think": 965,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": true,
     "probe_robust": true,
     "mimicry": false,
     "full": true,
     "gen_tokens": 2083,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1089,
     "n_think": 666,
     "forced": false
    }
   ]
  },
  {
   "task_id": 21,
   "family": "string",
   "depth": 3,
   "prompt_tokens": 204,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2215,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 919,
     "n_think": 768,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2214,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2303,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2329,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2237,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    }
   ]
  },
  {
   "task_id": 22,
   "family": "string",
   "depth": 3,
   "prompt_tokens": 171,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2562,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2158,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2093,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 936,
     "n_think": 685,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1658,
     "n_think": 1528,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    }
   ]
  },
  {
   "task_id": 23,
   "family": "string",
   "depth": 3,
   "prompt_tokens": 220,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2297,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2096,
     "n_think": 2028,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2133,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 869,
     "n_think": 828,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1054,
     "n_think": 960,
     "forced": false
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2125,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2139,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1082,
     "n_think": 1007,
     "forced": false
    }
   ]
  },
  {
   "task_id": 24,
   "family": "string",
   "depth": 3,
   "prompt_tokens": 173,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2214,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 729,
     "n_think": 709,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2486,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1007,
     "n_think": 987,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2068,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 731,
     "n_think": 667,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2068,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1366,
     "n_think": 1330,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 604,
     "n_think": 584,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": true,
     "probe_robust": false,
     "mimicry": true,
     "full": false,
     "gen_tokens": 2362,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2087,
     "n_think": 2048,
     "forced": true
    }
   ]
  },
  {
   "task_id": 25,
   "family": "string",
   "depth": 3,
   "prompt_tokens": 170,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2100,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 976,
     "n_think": 799,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1240,
     "n_think": 1087,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 721,
     "n_think": 607,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 876,
     "n_think": 762,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2094,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 845,
     "n_think": 751,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1142,
     "n_think": 1011,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2095,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 889,
     "n_think": 790,
     "forced": false
    }
   ]
  },
  {
   "task_id": 26,
   "family": "string",
   "depth": 3,
   "prompt_tokens": 189,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2127,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2068,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1004,
     "n_think": 903,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 724,
     "n_think": 646,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2100,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1056,
     "n_think": 947,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 766,
     "n_think": 697,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 724,
     "n_think": 658,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1056,
     "n_think": 1034,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1096,
     "n_think": 1013,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1261,
     "n_think": 1183,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2339,
     "n_think": 2048,
     "forced": true
    }
   ]
  },
  {
   "task_id": 27,
   "family": "string",
   "depth": 3,
   "prompt_tokens": 196,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2306,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2468,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2304,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2225,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2447,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1096,
     "n_think": 821,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2215,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1231,
     "n_think": 1036,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2372,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2242,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1259,
     "n_think": 959,
     "forced": false
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    }
   ]
  },
  {
   "task_id": 28,
   "family": "string",
   "depth": 3,
   "prompt_tokens": 174,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 1391,
    "n_think": 1245,
    "forced": false
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 654,
     "n_think": 624,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2081,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1429,
     "n_think": 1272,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1025,
     "n_think": 1000,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2068,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1529,
     "n_think": 1492,
     "forced": false
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2113,
     "n_think": 2048,
     "forced": true
    }
   ]
  },
  {
   "task_id": 29,
   "family": "string",
   "depth": 3,
   "prompt_tokens": 177,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 2562,
    "n_think": 2048,
    "forced": true
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2079,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 549,
     "n_think": 466,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2145,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2149,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    }
   ]
  },
  {
   "task_id": 30,
   "family": "string",
   "depth": 3,
   "prompt_tokens": 165,
   "greedy": {
    "parsed": true,
    "legacy": false,
    "probe_robust": false,
    "mimicry": false,
    "full": false,
    "gen_tokens": 1894,
    "n_think": 1853,
    "forced": false
   },
   "samples": [
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2136,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 1698,
     "n_think": 1652,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2090,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2094,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 885,
     "n_think": 735,
     "forced": false
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2115,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": false,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2562,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2085,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 2085,
     "n_think": 2048,
     "forced": true
    },
    {
     "parsed": true,
     "legacy": false,
     "probe_robust": false,
     "mimicry": false,
     "full": false,
     "gen_tokens": 930,
     "n_think": 789,
     "forced": false
    }
   ]
  }
 ]
}