[
  {
    "model": "llama3.1:8b",
    "run": 0,
    "score": 0.0,
    "tiers": {
      "tier1_core": 0.0,
      "tier2_splits": 0.0,
      "tier3_settle": 0.0,
      "tier4_validation": 0.0,
      "tier5_stress": 0.0
    },
    "steps": 7,
    "end_reason": "completed",
    "visible_pass_rate": 0.0,
    "full_solve": false,
    "wall_limit_minutes": 30
  },
  {
    "model": "llama3.1:8b",
    "run": 1,
    "score": 0.2603,
    "tiers": {
      "tier1_core": 0.3077,
      "tier2_splits": 0.0667,
      "tier3_settle": 0.0,
      "tier4_validation": 0.6667,
      "tier5_stress": 0.0
    },
    "steps": 40,
    "end_reason": "step_limit",
    "visible_pass_rate": 0.0,
    "full_solve": false,
    "wall_limit_minutes": 240
  },
  {
    "model": "llama3.1:8b",
    "run": 2,
    "score": 0.2276,
    "tiers": {
      "tier1_core": 0.0769,
      "tier2_splits": 0.0,
      "tier3_settle": 0.0,
      "tier4_validation": 0.8333,
      "tier5_stress": 0.0
    },
    "steps": 5,
    "end_reason": "completed",
    "visible_pass_rate": 0.0,
    "full_solve": false,
    "wall_limit_minutes": 30
  },
  {
    "model": "llama3.1:8b",
    "run": 3,
    "score": 0.1394,
    "tiers": {
      "tier1_core": 0.3077,
      "tier2_splits": 0.0,
      "tier3_settle": 0.0,
      "tier4_validation": 0.25,
      "tier5_stress": 0.0
    },
    "steps": 8,
    "end_reason": "completed",
    "visible_pass_rate": 0.0,
    "full_solve": false,
    "wall_limit_minutes": 30
  },
  {
    "model": "llama3.1:8b",
    "run": 4,
    "score": 0.0,
    "tiers": {
      "tier1_core": 0.0,
      "tier2_splits": 0.0,
      "tier3_settle": 0.0,
      "tier4_validation": 0.0,
      "tier5_stress": 0.0
    },
    "steps": 40,
    "end_reason": "step_limit",
    "visible_pass_rate": 0.0,
    "full_solve": false,
    "wall_limit_minutes": 30
  },
  {
    "model": "qwen2.5-coder:14b",
    "run": 0,
    "score": 0.4309,
    "tiers": {
      "tier1_core": 0.6154,
      "tier2_splits": 0.0667,
      "tier3_settle": 0.3333,
      "tier4_validation": 0.7083,
      "tier5_stress": 0.0
    },
    "steps": 40,
    "end_reason": "step_limit",
    "visible_pass_rate": 0.3333,
    "full_solve": false,
    "wall_limit_minutes": 240
  },
  {
    "model": "qwen2.5-coder:14b",
    "run": 1,
    "score": 0.2091,
    "tiers": {
      "tier1_core": 0.4615,
      "tier2_splits": 0.0,
      "tier3_settle": 0.0,
      "tier4_validation": 0.375,
      "tier5_stress": 0.0
    },
    "steps": 40,
    "end_reason": "step_limit",
    "visible_pass_rate": 0.3333,
    "full_solve": false,
    "wall_limit_minutes": 240
  },
  {
    "model": "qwen2.5-coder:14b",
    "run": 2,
    "score": 0.7058,
    "tiers": {
      "tier1_core": 0.9231,
      "tier2_splits": 0.4,
      "tier3_settle": 0.5,
      "tier4_validation": 1.0,
      "tier5_stress": 0.0333
    },
    "steps": 40,
    "end_reason": "step_limit",
    "visible_pass_rate": 0.6667,
    "full_solve": false,
    "wall_limit_minutes": 240
  },
  {
    "model": "qwen2.5-coder:14b",
    "run": 3,
    "score": 0.3284,
    "tiers": {
      "tier1_core": 0.5385,
      "tier2_splits": 0.0667,
      "tier3_settle": 0.0,
      "tier4_validation": 0.7083,
      "tier5_stress": 0.0
    },
    "steps": 40,
    "end_reason": "step_limit",
    "visible_pass_rate": 0.3333,
    "full_solve": false,
    "wall_limit_minutes": 240
  },
  {
    "model": "qwen2.5-coder:14b",
    "run": 4,
    "score": 0.3909,
    "tiers": {
      "tier1_core": 0.5385,
      "tier2_splits": 0.0667,
      "tier3_settle": 0.0,
      "tier4_validation": 0.9583,
      "tier5_stress": 0.0
    },
    "steps": 40,
    "end_reason": "step_limit",
    "visible_pass_rate": 0.3333,
    "full_solve": false,
    "wall_limit_minutes": 240
  },
  {
    "model": "qwen2.5-coder:7b",
    "run": 0,
    "score": 0.3716,
    "tiers": {
      "tier1_core": 0.4615,
      "tier2_splits": 0.0667,
      "tier3_settle": 0.0,
      "tier4_validation": 0.9583,
      "tier5_stress": 0.0
    },
    "steps": 3,
    "end_reason": "completed",
    "visible_pass_rate": 0.3333,
    "full_solve": false,
    "wall_limit_minutes": 10
  },
  {
    "model": "qwen2.5-coder:7b",
    "run": 1,
    "score": 0.2644,
    "tiers": {
      "tier1_core": 0.3077,
      "tier2_splits": 0.0,
      "tier3_settle": 0.0,
      "tier4_validation": 0.75,
      "tier5_stress": 0.0
    },
    "steps": 40,
    "end_reason": "step_limit",
    "visible_pass_rate": 0.3333,
    "full_solve": false,
    "wall_limit_minutes": 240
  },
  {
    "model": "qwen2.5-coder:7b",
    "run": 2,
    "score": 0.3149,
    "tiers": {
      "tier1_core": 0.3846,
      "tier2_splits": 0.0,
      "tier3_settle": 0.0,
      "tier4_validation": 0.875,
      "tier5_stress": 0.0
    },
    "steps": 40,
    "end_reason": "step_limit",
    "visible_pass_rate": 0.0,
    "full_solve": false,
    "wall_limit_minutes": 240
  },
  {
    "model": "qwen2.5-coder:7b",
    "run": 3,
    "score": 0.2869,
    "tiers": {
      "tier1_core": 0.2308,
      "tier2_splits": 0.0,
      "tier3_settle": 0.0,
      "tier4_validation": 0.9167,
      "tier5_stress": 0.0
    },
    "steps": 3,
    "end_reason": "completed",
    "visible_pass_rate": 0.0,
    "full_solve": false,
    "wall_limit_minutes": 30
  },
  {
    "model": "qwen2.5-coder:7b",
    "run": 4,
    "score": 0.3061,
    "tiers": {
      "tier1_core": 0.3077,
      "tier2_splits": 0.0,
      "tier3_settle": 0.0,
      "tier4_validation": 0.9167,
      "tier5_stress": 0.0
    },
    "steps": 40,
    "end_reason": "step_limit",
    "visible_pass_rate": 0.0,
    "full_solve": false,
    "wall_limit_minutes": 240
  },
  {
    "model": "qwen3:14b",
    "run": 0,
    "score": 0.7562,
    "tiers": {
      "tier1_core": 1.0,
      "tier2_splits": 0.4,
      "tier3_settle": 0.6667,
      "tier4_validation": 0.9583,
      "tier5_stress": 0.0333
    },
    "steps": 4,
    "end_reason": "completed",
    "visible_pass_rate": 0.6667,
    "full_solve": false,
    "wall_limit_minutes": 30
  },
  {
    "model": "qwen3:14b",
    "run": 1,
    "score": 0.7312,
    "tiers": {
      "tier1_core": 1.0,
      "tier2_splits": 0.4667,
      "tier3_settle": 0.5,
      "tier4_validation": 0.9583,
      "tier5_stress": 0.0667
    },
    "steps": 3,
    "end_reason": "completed",
    "visible_pass_rate": 0.6667,
    "full_solve": false,
    "wall_limit_minutes": 30
  },
  {
    "model": "qwen3:14b",
    "run": 2,
    "score": 0.4856,
    "tiers": {
      "tier1_core": 0.6923,
      "tier2_splits": 0.3333,
      "tier3_settle": 0.0,
      "tier4_validation": 0.9167,
      "tier5_stress": 0.0
    },
    "steps": 6,
    "end_reason": "completed",
    "visible_pass_rate": 0.0,
    "full_solve": false,
    "wall_limit_minutes": 30
  },
  {
    "model": "qwen3:14b",
    "run": 3,
    "score": 0.7162,
    "tiers": {
      "tier1_core": 0.9231,
      "tier2_splits": 0.4,
      "tier3_settle": 0.8333,
      "tier4_validation": 0.7083,
      "tier5_stress": 0.0333
    },
    "steps": 4,
    "end_reason": "completed",
    "visible_pass_rate": 0.6667,
    "full_solve": false,
    "wall_limit_minutes": 30
  },
  {
    "model": "qwen3:14b",
    "run": 4,
    "score": 0.8287,
    "tiers": {
      "tier1_core": 0.9231,
      "tier2_splits": 0.6,
      "tier3_settle": 0.8333,
      "tier4_validation": 0.9583,
      "tier5_stress": 0.0333
    },
    "steps": 14,
    "end_reason": "completed",
    "visible_pass_rate": 1.0,
    "full_solve": false,
    "wall_limit_minutes": 240
  }
]
