{
  "campaign": "2026-09-09",
  "models": [
    "qwen2.5-7b-instruct",
    "qwen2.5-coder-14b-instruct"
  ],
  "tasks": [
    "coop-scheduler",
    "frame-corrupt",
    "mqtt-device",
    "noisy-frames",
    "p-regulator",
    "pid-antiwindup",
    "protocol",
    "protocol-retry",
    "system-id",
    "traffic-light",
    "uart-echo",
    "uart-menu",
    "watchdog"
  ],
  "pass_at_k": 0.31,
  "groups": [
    {
      "model": "qwen2.5-7b-instruct",
      "task": "coop-scheduler",
      "attempts": 2,
      "solved": 0,
      "total_iterations": 10,
      "total_duration": 452.87,
      "success_rate": 0.0,
      "avg_iterations": 5.0,
      "avg_duration": 226.4,
      "passed": false,
      "tags": [
        "control"
      ],
      "level": 3
    },
    {
      "model": "qwen2.5-7b-instruct",
      "task": "frame-corrupt",
      "attempts": 2,
      "solved": 0,
      "total_iterations": 10,
      "total_duration": 326.78,
      "success_rate": 0.0,
      "avg_iterations": 5.0,
      "avg_duration": 163.4,
      "passed": false,
      "tags": [
        "resilience"
      ],
      "level": 4
    },
    {
      "model": "qwen2.5-7b-instruct",
      "task": "mqtt-device",
      "attempts": 2,
      "solved": 0,
      "total_iterations": 10,
      "total_duration": 148.44,
      "success_rate": 0.0,
      "avg_iterations": 5.0,
      "avg_duration": 74.2,
      "passed": false,
      "tags": [
        "protocol"
      ],
      "level": 5
    },
    {
      "model": "qwen2.5-7b-instruct",
      "task": "noisy-frames",
      "attempts": 2,
      "solved": 0,
      "total_iterations": 10,
      "total_duration": 214.97,
      "success_rate": 0.0,
      "avg_iterations": 5.0,
      "avg_duration": 107.5,
      "passed": false,
      "tags": [
        "resilience",
        "protocol"
      ],
      "level": 3
    },
    {
      "model": "qwen2.5-7b-instruct",
      "task": "p-regulator",
      "attempts": 2,
      "solved": 2,
      "total_iterations": 2,
      "total_duration": 6.54,
      "success_rate": 1.0,
      "avg_iterations": 1.0,
      "avg_duration": 3.3,
      "passed": true,
      "tags": [
        "control"
      ],
      "level": 2
    },
    {
      "model": "qwen2.5-7b-instruct",
      "task": "pid-antiwindup",
      "attempts": 2,
      "solved": 0,
      "total_iterations": 10,
      "total_duration": 53.099999999999994,
      "success_rate": 0.0,
      "avg_iterations": 5.0,
      "avg_duration": 26.5,
      "passed": false,
      "tags": [
        "control"
      ],
      "level": 3
    },
    {
      "model": "qwen2.5-7b-instruct",
      "task": "protocol",
      "attempts": 2,
      "solved": 0,
      "total_iterations": 10,
      "total_duration": 53.83,
      "success_rate": 0.0,
      "avg_iterations": 5.0,
      "avg_duration": 26.9,
      "passed": false,
      "tags": [
        "protocol"
      ],
      "level": 2
    },
    {
      "model": "qwen2.5-7b-instruct",
      "task": "protocol-retry",
      "attempts": 2,
      "solved": 0,
      "total_iterations": 10,
      "total_duration": 59.17,
      "success_rate": 0.0,
      "avg_iterations": 5.0,
      "avg_duration": 29.6,
      "passed": false,
      "tags": [
        "protocol",
        "resilience"
      ],
      "level": 3
    },
    {
      "model": "qwen2.5-7b-instruct",
      "task": "system-id",
      "attempts": 2,
      "solved": 0,
      "total_iterations": 10,
      "total_duration": 52.72,
      "success_rate": 0.0,
      "avg_iterations": 5.0,
      "avg_duration": 26.4,
      "passed": false,
      "tags": [
        "control"
      ],
      "level": 4
    },
    {
      "model": "qwen2.5-7b-instruct",
      "task": "traffic-light",
      "attempts": 2,
      "solved": 0,
      "total_iterations": 10,
      "total_duration": 40.980000000000004,
      "success_rate": 0.0,
      "avg_iterations": 5.0,
      "avg_duration": 20.5,
      "passed": false,
      "tags": [
        "fsm"
      ],
      "level": 3
    },
    {
      "model": "qwen2.5-7b-instruct",
      "task": "uart-echo",
      "attempts": 2,
      "solved": 2,
      "total_iterations": 3,
      "total_duration": 22.55,
      "success_rate": 1.0,
      "avg_iterations": 1.5,
      "avg_duration": 11.3,
      "passed": true,
      "tags": [
        "io"
      ],
      "level": 1
    },
    {
      "model": "qwen2.5-7b-instruct",
      "task": "uart-menu",
      "attempts": 2,
      "solved": 2,
      "total_iterations": 5,
      "total_duration": 188.84,
      "success_rate": 1.0,
      "avg_iterations": 2.5,
      "avg_duration": 94.4,
      "passed": true,
      "tags": [
        "fsm"
      ],
      "level": 3
    },
    {
      "model": "qwen2.5-7b-instruct",
      "task": "watchdog",
      "attempts": 2,
      "solved": 2,
      "total_iterations": 2,
      "total_duration": 9.65,
      "success_rate": 1.0,
      "avg_iterations": 1.0,
      "avg_duration": 4.8,
      "passed": true,
      "tags": [
        "resilience"
      ],
      "level": 3
    },
    {
      "model": "qwen2.5-coder-14b-instruct",
      "task": "coop-scheduler",
      "attempts": 2,
      "solved": 0,
      "total_iterations": 10,
      "total_duration": 950.96,
      "success_rate": 0.0,
      "avg_iterations": 5.0,
      "avg_duration": 475.5,
      "passed": false,
      "tags": [
        "control"
      ],
      "level": 3
    },
    {
      "model": "qwen2.5-coder-14b-instruct",
      "task": "frame-corrupt",
      "attempts": 2,
      "solved": 0,
      "total_iterations": 10,
      "total_duration": 840.1700000000001,
      "success_rate": 0.0,
      "avg_iterations": 5.0,
      "avg_duration": 420.1,
      "passed": false,
      "tags": [
        "resilience"
      ],
      "level": 4
    },
    {
      "model": "qwen2.5-coder-14b-instruct",
      "task": "mqtt-device",
      "attempts": 2,
      "solved": 0,
      "total_iterations": 10,
      "total_duration": 2042.6,
      "success_rate": 0.0,
      "avg_iterations": 5.0,
      "avg_duration": 1021.3,
      "passed": false,
      "tags": [
        "protocol"
      ],
      "level": 5
    },
    {
      "model": "qwen2.5-coder-14b-instruct",
      "task": "noisy-frames",
      "attempts": 2,
      "solved": 0,
      "total_iterations": 10,
      "total_duration": 701.05,
      "success_rate": 0.0,
      "avg_iterations": 5.0,
      "avg_duration": 350.5,
      "passed": false,
      "tags": [
        "resilience",
        "protocol"
      ],
      "level": 3
    },
    {
      "model": "qwen2.5-coder-14b-instruct",
      "task": "p-regulator",
      "attempts": 2,
      "solved": 1,
      "total_iterations": 6,
      "total_duration": 137.53,
      "success_rate": 0.5,
      "avg_iterations": 3.0,
      "avg_duration": 68.8,
      "passed": true,
      "tags": [
        "control"
      ],
      "level": 2
    },
    {
      "model": "qwen2.5-coder-14b-instruct",
      "task": "pid-antiwindup",
      "attempts": 2,
      "solved": 0,
      "total_iterations": 10,
      "total_duration": 448.35,
      "success_rate": 0.0,
      "avg_iterations": 5.0,
      "avg_duration": 224.2,
      "passed": false,
      "tags": [
        "control"
      ],
      "level": 3
    },
    {
      "model": "qwen2.5-coder-14b-instruct",
      "task": "protocol",
      "attempts": 2,
      "solved": 0,
      "total_iterations": 10,
      "total_duration": 367.95000000000005,
      "success_rate": 0.0,
      "avg_iterations": 5.0,
      "avg_duration": 184.0,
      "passed": false,
      "tags": [
        "protocol"
      ],
      "level": 2
    },
    {
      "model": "qwen2.5-coder-14b-instruct",
      "task": "protocol-retry",
      "attempts": 2,
      "solved": 0,
      "total_iterations": 10,
      "total_duration": 584.33,
      "success_rate": 0.0,
      "avg_iterations": 5.0,
      "avg_duration": 292.2,
      "passed": false,
      "tags": [
        "protocol",
        "resilience"
      ],
      "level": 3
    },
    {
      "model": "qwen2.5-coder-14b-instruct",
      "task": "system-id",
      "attempts": 2,
      "solved": 0,
      "total_iterations": 10,
      "total_duration": 958.26,
      "success_rate": 0.0,
      "avg_iterations": 5.0,
      "avg_duration": 479.1,
      "passed": false,
      "tags": [
        "control"
      ],
      "level": 4
    },
    {
      "model": "qwen2.5-coder-14b-instruct",
      "task": "traffic-light",
      "attempts": 2,
      "solved": 0,
      "total_iterations": 10,
      "total_duration": 694.41,
      "success_rate": 0.0,
      "avg_iterations": 5.0,
      "avg_duration": 347.2,
      "passed": false,
      "tags": [
        "fsm"
      ],
      "level": 3
    },
    {
      "model": "qwen2.5-coder-14b-instruct",
      "task": "uart-echo",
      "attempts": 2,
      "solved": 2,
      "total_iterations": 2,
      "total_duration": 25.25,
      "success_rate": 1.0,
      "avg_iterations": 1.0,
      "avg_duration": 12.6,
      "passed": true,
      "tags": [
        "io"
      ],
      "level": 1
    },
    {
      "model": "qwen2.5-coder-14b-instruct",
      "task": "uart-menu",
      "attempts": 2,
      "solved": 2,
      "total_iterations": 2,
      "total_duration": 85.59,
      "success_rate": 1.0,
      "avg_iterations": 1.0,
      "avg_duration": 42.8,
      "passed": true,
      "tags": [
        "fsm"
      ],
      "level": 3
    },
    {
      "model": "qwen2.5-coder-14b-instruct",
      "task": "watchdog",
      "attempts": 2,
      "solved": 2,
      "total_iterations": 5,
      "total_duration": 362.90000000000003,
      "success_rate": 1.0,
      "avg_iterations": 2.5,
      "avg_duration": 181.5,
      "passed": true,
      "tags": [
        "resilience"
      ],
      "level": 3
    }
  ],
  "class_profile": {
    "qwen2.5-7b-instruct": {
      "control": {
        "attempts": 8,
        "solved": 2,
        "success_rate": 0.25
      },
      "fsm": {
        "attempts": 4,
        "solved": 2,
        "success_rate": 0.5
      },
      "io": {
        "attempts": 2,
        "solved": 2,
        "success_rate": 1.0
      },
      "protocol": {
        "attempts": 8,
        "solved": 0,
        "success_rate": 0.0
      },
      "resilience": {
        "attempts": 8,
        "solved": 2,
        "success_rate": 0.25
      }
    },
    "qwen2.5-coder-14b-instruct": {
      "control": {
        "attempts": 8,
        "solved": 1,
        "success_rate": 0.12
      },
      "fsm": {
        "attempts": 4,
        "solved": 2,
        "success_rate": 0.5
      },
      "io": {
        "attempts": 2,
        "solved": 2,
        "success_rate": 1.0
      },
      "protocol": {
        "attempts": 8,
        "solved": 0,
        "success_rate": 0.0
      },
      "resilience": {
        "attempts": 8,
        "solved": 2,
        "success_rate": 0.25
      }
    }
  }
}