{
  "bench_version": 1,
  "track": "conditional_ot_domains",
  "model": "shared_mlp_tiny",
  "metric_direction": "lower is better",
  "n_seeds": 8,
  "baseline": {
    "best_cfg": {
      "lr": 0.006,
      "eta": 0.15
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.003,
          "eta": 0.15
        },
        "mean": 0.07499999739229679
      },
      {
        "cfg": {
          "lr": 0.003,
          "eta": 0.35
        },
        "mean": 0.07499999739229679
      },
      {
        "cfg": {
          "lr": 0.003,
          "eta": 0.6
        },
        "mean": 0.07499999739229679
      },
      {
        "cfg": {
          "lr": 0.006,
          "eta": 0.15
        },
        "mean": 0.07249999791383743
      },
      {
        "cfg": {
          "lr": 0.006,
          "eta": 0.35
        },
        "mean": 0.07249999791383743
      },
      {
        "cfg": {
          "lr": 0.006,
          "eta": 0.6
        },
        "mean": 0.07249999791383743
      },
      {
        "cfg": {
          "lr": 0.012,
          "eta": 0.15
        },
        "mean": 0.08874999731779099
      },
      {
        "cfg": {
          "lr": 0.012,
          "eta": 0.35
        },
        "mean": 0.08874999731779099
      },
      {
        "cfg": {
          "lr": 0.012,
          "eta": 0.6
        },
        "mean": 0.08874999731779099
      }
    ],
    "full": {
      "mean": 0.07749999780207872,
      "std": 0.010825316902066397,
      "per_seed": [
        0.07750000059604645,
        0.07249999791383743,
        0.07499999552965164,
        0.06499999761581421,
        0.08249999582767487,
        0.0625,
        0.08749999850988388,
        0.09749999642372131
      ],
      "n": 8
    }
  },
  "idea": {
    "mean": 0.07249999837949872,
    "std": 0.012437343083642742,
    "per_seed": [
      0.08500000089406967,
      0.06749999523162842,
      0.08249999582767487,
      0.05249999836087227,
      0.054999999701976776,
      0.07750000059604645,
      0.07249999791383743,
      0.08749999850988388
    ],
    "n": 8
  },
  "comparison": {
    "delta_mean": -0.004999999422580004,
    "idea_wins": 5,
    "n_pairs": 8,
    "per_seed_diffs": [
      0.007500000298023224,
      -0.005000002682209015,
      0.007500000298023224,
      -0.01249999925494194,
      -0.02749999612569809,
      0.015000000596046448,
      -0.015000000596046448,
      -0.009999997913837433
    ],
    "p_value": 0.39075,
    "mde": 0.011824999342373263,
    "mde_rel_pct": 15.258064100301292,
    "verdict": "no significant win",
    "system_worked": false
  },
  "mechanism_signature": {
    "mechanism_signature": {
      "prediction": "three-source latent consensus has approximately one-third of pooled feature variance",
      "predicted": 0.2850805396835009,
      "observed": 0.26271023021803963,
      "observed_reduction_ratio": 0.3066594176419761,
      "confirmed": true,
      "n_behavior_models": 8
    },
    "idea_sweep": [
      {
        "cfg": {
          "lr": 0.003,
          "eta": 0.15
        },
        "result": {
          "mean": 0.07593749789521098,
          "std": 0.016720940857393073,
          "per_seed": [
            0.07999999821186066,
            0.07499999552965164,
            0.10249999910593033,
            0.044999998062849045,
            0.057499997317790985,
            0.08749999850988388,
            0.07499999552965164,
            0.08500000089406967
          ],
          "n": 8
        }
      },
      {
        "cfg": {
          "lr": 0.003,
          "eta": 0.35
        },
        "result": {
          "mean": 0.07593749789521098,
          "std": 0.016720940857393073,
          "per_seed": [
            0.07999999821186066,
            0.07499999552965164,
            0.10249999910593033,
            0.044999998062849045,
            0.057499997317790985,
            0.08749999850988388,
            0.07499999552965164,
            0.08500000089406967
          ],
          "n": 8
        }
      },
      {
        "cfg": {
          "lr": 0.003,
          "eta": 0.6
        },
        "result": {
          "mean": 0.07531249849125743,
          "std": 0.01688367871052282,
          "per_seed": [
            0.07999999821186066,
            0.07499999552965164,
            0.10249999910593033,
            0.044999998062849045,
            0.054999999701976776,
            0.08500000089406967,
            0.07499999552965164,
            0.08500000089406967
          ],
          "n": 8
        }
      },
      {
        "cfg": {
          "lr": 0.006,
          "eta": 0.15
        },
        "result": {
          "mean": 0.07499999832361937,
          "std": 0.012990381393724975,
          "per_seed": [
            0.08749999850988388,
            0.07499999552965164,
            0.07499999552965164,
            0.04749999940395355,
            0.07999999821186066,
            0.06499999761581421,
            0.07750000059604645,
            0.0925000011920929
          ],
          "n": 8
        }
      },
      {
        "cfg": {
          "lr": 0.006,
          "eta": 0.35
        },
        "result": {
          "mean": 0.07749999780207872,
          "std": 0.014361406700648869,
          "per_seed": [
            0.08249999582767487,
            0.07499999552965164,
            0.08249999582767487,
            0.04749999940395355,
            0.0925000011920929,
            0.06499999761581421,
            0.07999999821186066,
            0.0949999988079071
          ],
          "n": 8
        }
      },
      {
        "cfg": {
          "lr": 0.006,
          "eta": 0.6
        },
        "result": {
          "mean": 0.07749999780207872,
          "std": 0.01499999950950353,
          "per_seed": [
            0.08249999582767487,
            0.07499999552965164,
            0.07750000059604645,
            0.04749999940395355,
            0.09749999642372131,
            0.06499999761581421,
            0.07999999821186066,
            0.0949999988079071
          ],
          "n": 8
        }
      },
      {
        "cfg": {
          "lr": 0.012,
          "eta": 0.15
        },
        "result": {
          "mean": 0.07249999837949872,
          "std": 0.012437343083642742,
          "per_seed": [
            0.08500000089406967,
            0.06749999523162842,
            0.08249999582767487,
            0.05249999836087227,
            0.054999999701976776,
            0.07750000059604645,
            0.07249999791383743,
            0.08749999850988388
          ],
          "n": 8
        }
      },
      {
        "cfg": {
          "lr": 0.012,
          "eta": 0.35
        },
        "result": {
          "mean": 0.07562499865889549,
          "std": 0.013506364375069949,
          "per_seed": [
            0.0949999988079071,
            0.07249999791383743,
            0.07249999791383743,
            0.05249999836087227,
            0.05999999865889549,
            0.08749999850988388,
            0.07750000059604645,
            0.08749999850988388
          ],
          "n": 8
        }
      },
      {
        "cfg": {
          "lr": 0.012,
          "eta": 0.6
        },
        "result": {
          "mean": 0.0796874975785613,
          "std": 0.01454720197649073,
          "per_seed": [
            0.0949999988079071,
            0.07249999791383743,
            0.07000000029802322,
            0.04999999701976776,
            0.07999999821186066,
            0.09749999642372131,
            0.08249999582767487,
            0.08999999612569809
          ],
          "n": 8
        }
      }
    ],
    "custom_track": {
      "name": "conditional_ot_domains",
      "file": "conditional_ot_track.py",
      "domain": "domain_generalization"
    },
    "protocol": {
      "paired_seeds": [
        0,
        1,
        2,
        3,
        4,
        5,
        6,
        7
      ],
      "sweep_seeds": [
        0,
        1,
        2,
        3
      ],
      "grid_union": [
        {
          "lr": 0.003,
          "eta": 0.15
        },
        {
          "lr": 0.003,
          "eta": 0.35
        },
        {
          "lr": 0.003,
          "eta": 0.6
        },
        {
          "lr": 0.006,
          "eta": 0.15
        },
        {
          "lr": 0.006,
          "eta": 0.35
        },
        {
          "lr": 0.006,
          "eta": 0.6
        },
        {
          "lr": 0.012,
          "eta": 0.15
        },
        {
          "lr": 0.012,
          "eta": 0.35
        },
        {
          "lr": 0.012,
          "eta": 0.6
        }
      ],
      "baseline_tuned_and_reevaluated": true,
      "task_match": "multi-source domain generalization with latent feature consensus"
    }
  }
}