{
  "bench_version": 1,
  "track": "expert_balanced_regression",
  "model": "top1_moe_shared",
  "metric_direction": "lower is better",
  "n_seeds": 8,
  "baseline": {
    "best_cfg": {
      "lr": 0.01,
      "epochs": 12,
      "rho": 0.5
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.001,
          "epochs": 12,
          "rho": 0.5
        },
        "mean": 0.5207640379667282
      },
      {
        "cfg": {
          "lr": 0.001,
          "epochs": 12,
          "rho": 1.0
        },
        "mean": 0.5207640379667282
      },
      {
        "cfg": {
          "lr": 0.003,
          "epochs": 12,
          "rho": 0.5
        },
        "mean": 0.3618839532136917
      },
      {
        "cfg": {
          "lr": 0.003,
          "epochs": 12,
          "rho": 1.0
        },
        "mean": 0.3618839532136917
      },
      {
        "cfg": {
          "lr": 0.01,
          "epochs": 12,
          "rho": 0.5
        },
        "mean": 0.21788224577903748
      },
      {
        "cfg": {
          "lr": 0.01,
          "epochs": 12,
          "rho": 1.0
        },
        "mean": 0.21788224577903748
      }
    ],
    "full": {
      "mean": 0.21290172263979912,
      "std": 0.02201949504273137,
      "per_seed": [
        0.23030464351177216,
        0.19820281863212585,
        0.18614302575588226,
        0.25687849521636963,
        0.21424928307533264,
        0.18679030239582062,
        0.21802611649036407,
        0.2126190960407257
      ],
      "n": 8
    }
  },
  "idea": {
    "mean": 0.2056338880211115,
    "std": 0.020574444566337256,
    "per_seed": [
      0.2081810086965561,
      0.20322394371032715,
      0.18413016200065613,
      0.24816179275512695,
      0.19936268031597137,
      0.17569208145141602,
      0.20796272158622742,
      0.21835671365261078
    ],
    "n": 8,
    "config": {
      "lr": 0.01,
      "epochs": 12,
      "rho": 1.0
    }
  },
  "comparison": {
    "delta_mean": -0.00726783461868763,
    "idea_wins": 6,
    "n_pairs": 8,
    "per_seed_diffs": [
      -0.022123634815216064,
      0.005021125078201294,
      -0.0020128637552261353,
      -0.008716702461242676,
      -0.014886602759361267,
      -0.011098220944404602,
      -0.010063394904136658,
      0.005737617611885071
    ],
    "p_value": 0.08055,
    "mde": 0.008064204749019211,
    "mde_rel_pct": 3.7877592764539316,
    "verdict": "no significant win",
    "system_worked": false
  },
  "mechanism_signature": {
    "prediction": "trained VDC assignments maintain max leaf prefix discrepancy <=1",
    "predicted_max_discrepancy": 1.0,
    "observed_mean_max_discrepancy": 0.875,
    "observed_final_load_variance": 0.0,
    "confirmed": true
  },
  "custom_track": {
    "name": "expert_balanced_regression",
    "file": "expert_balanced_regression.py",
    "domain": "moe-routing"
  }
}