{
  "bench_version": 1,
  "track": "router_regime_regression",
  "model": "custom_mlp_moe",
  "metric_direction": "lower is better",
  "n_seeds": 8,
  "baseline": {
    "best_cfg": {
      "lr": 0.012,
      "temp": 0.7
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.003,
          "temp": 0.7
        },
        "mean": 0.5712137892842293
      },
      {
        "cfg": {
          "lr": 0.003,
          "temp": 1.0
        },
        "mean": 0.6234312355518341
      },
      {
        "cfg": {
          "lr": 0.006,
          "temp": 0.7
        },
        "mean": 0.4075416550040245
      },
      {
        "cfg": {
          "lr": 0.006,
          "temp": 1.0
        },
        "mean": 0.44819550961256027
      },
      {
        "cfg": {
          "lr": 0.012,
          "temp": 0.7
        },
        "mean": 0.27756691724061966
      },
      {
        "cfg": {
          "lr": 0.012,
          "temp": 1.0
        },
        "mean": 0.31893689185380936
      }
    ],
    "full": {
      "mean": 0.3102635517716408,
      "std": 0.07497470190378723,
      "per_seed": [
        0.30197641253471375,
        0.24877610802650452,
        0.29255229234695435,
        0.26696285605430603,
        0.28872963786125183,
        0.5001022219657898,
        0.26213279366493225,
        0.3208760917186737
      ],
      "n": 8
    }
  },
  "idea": {
    "cfg": {
      "lr": 0.012,
      "temp": 0.7,
      "rho": 0.3
    },
    "per_seed": [
      0.3346004784107208,
      0.27418217062950134,
      0.3511717915534973,
      0.32937872409820557,
      0.3155486285686493,
      0.5077916979789734,
      0.2751004695892334,
      0.34620028734207153
    ],
    "mean": 0.3417467810213566,
    "std": 0.07326728763277315,
    "signature_summary": {
      "mean_observed_cap": 0.2041564326380133,
      "mean_predicted_cap": 0.2041564326380133
    }
  },
  "comparison": {
    "delta_mean": 0.031483229249715805,
    "idea_wins": 0,
    "n_pairs": 8,
    "per_seed_diffs": [
      0.03262406587600708,
      0.025406062602996826,
      0.05861949920654297,
      0.062415868043899536,
      0.02681899070739746,
      0.007689476013183594,
      0.012967675924301147,
      0.025324195623397827
    ],
    "p_value": 0.0081,
    "mde": 0.01642202588917077,
    "mde_rel_pct": 5.292927833578616,
    "verdict": "idea worse (significant)",
    "system_worked": false
  },
  "mechanism_signature": {
    "prediction": "resistance-scaled natural-gradient router updates obey the pairwise cap on trained NN updates",
    "predicted_vs_observed": {
      "mean_observed_cap": 0.2041564326380133,
      "mean_predicted_cap": 0.2041564326380133
    },
    "confirmed": true,
    "math_sanity": {
      "max_resistance_ratio": 0.9998409991250671,
      "min_covariance_bound_eigenvalue": -1.1862584540283742e-16,
      "trust_cap_max_abs_error": 2.7755575615628914e-17,
      "passed": true
    }
  },
  "custom_track": {
    "name": "router_regime_regression",
    "file": "custom_router_track.py",
    "domain": "moe-routing"
  },
  "idea_sweep": [
    {
      "cfg": {
        "lr": 0.012,
        "temp": 0.7,
        "rho": 0.15
      },
      "per_seed": [
        0.33006882667541504,
        0.27459344267845154,
        0.3564022481441498,
        0.33827874064445496,
        0.31755706667900085,
        0.5151143074035645,
        0.2753690779209137,
        0.3360668420791626
      ],
      "mean": 0.3429313190281391,
      "std": 0.07553734469571749,
      "signature_summary": {
        "mean_observed_cap": 0.13508778517103004,
        "mean_predicted_cap": 0.13508778517103004
      }
    },
    {
      "cfg": {
        "lr": 0.012,
        "temp": 0.7,
        "rho": 0.3
      },
      "per_seed": [
        0.3346004784107208,
        0.27418217062950134,
        0.3511717915534973,
        0.32937872409820557,
        0.3155486285686493,
        0.5077916979789734,
        0.2751004695892334,
        0.34620028734207153
      ],
      "mean": 0.3417467810213566,
      "std": 0.07326728763277315,
      "signature_summary": {
        "mean_observed_cap": 0.2041564326380133,
        "mean_predicted_cap": 0.2041564326380133
      }
    },
    {
      "cfg": {
        "lr": 0.012,
        "temp": 0.7,
        "rho": 0.6
      },
      "per_seed": [
        0.34508222341537476,
        0.2762308120727539,
        0.3442482650279999,
        0.323551744222641,
        0.31441816687583923,
        0.5117174983024597,
        0.2795194685459137,
        0.36291345953941345
      ],
      "mean": 0.34471020475029945,
      "std": 0.07418865304996113,
      "signature_summary": {
        "mean_observed_cap": 0.23960371281639814,
        "mean_predicted_cap": 0.23960371281639814
      }
    },
    {
      "cfg": {
        "lr": 0.003,
        "temp": 0.7,
        "rho": 0.3
      },
      "per_seed": [
        0.6505195498466492,
        0.5359265208244324,
        0.6745683550834656,
        0.5784613490104675,
        0.6141240000724792,
        0.7810488343238831,
        0.5857309103012085,
        0.6731484532356262
      ],
      "mean": 0.6366909965872765,
      "std": 0.07601186541945128,
      "signature_summary": {
        "mean_observed_cap": 0.17640141463740155,
        "mean_predicted_cap": 0.17640141463740155
      }
    },
    {
      "cfg": {
        "lr": 0.006,
        "temp": 0.7,
        "rho": 0.3
      },
      "per_seed": [
        0.49773088097572327,
        0.380596786737442,
        0.49826493859291077,
        0.43299683928489685,
        0.4217953383922577,
        0.6373190879821777,
        0.4525613486766815,
        0.5051671862602234
      ],
      "mean": 0.47830405086278915,
      "std": 0.07767939883669912,
      "signature_summary": {
        "mean_observed_cap": 0.19019251237068713,
        "mean_predicted_cap": 0.19019251237068713
      }
    }
  ]
}