{
  "bench_version": 1,
  "track": "unordered_pointset_denoising",
  "model": "mlp_tiny",
  "metric_direction": "lower is better",
  "n_seeds": 8,
  "baseline": {
    "best_cfg": {
      "lr": 0.003,
      "weight_decay": 0.0001
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.001,
          "weight_decay": 0.0
        },
        "mean": 0.09564068168401718
      },
      {
        "cfg": {
          "lr": 0.001,
          "weight_decay": 0.0001
        },
        "mean": 0.09560250677168369
      },
      {
        "cfg": {
          "lr": 0.002,
          "weight_decay": 0.0
        },
        "mean": 0.084096509963274
      },
      {
        "cfg": {
          "lr": 0.002,
          "weight_decay": 0.0001
        },
        "mean": 0.0841062506660819
      },
      {
        "cfg": {
          "lr": 0.003,
          "weight_decay": 0.0
        },
        "mean": 0.07198812533169985
      },
      {
        "cfg": {
          "lr": 0.003,
          "weight_decay": 0.0001
        },
        "mean": 0.07177905738353729
      }
    ],
    "full": {
      "mean": 0.07177905738353729,
      "std": 0.0038503444468855823,
      "per_seed": [
        0.07469823211431503,
        0.06881649047136307,
        0.06867200881242752,
        0.0654606819152832,
        0.0715550109744072,
        0.07852364331483841,
        0.07229366153478622,
        0.07421272993087769
      ],
      "n": 8,
      "collision_per_seed": [
        0.054999999701976776,
        0.042500000447034836,
        0.05374999716877937,
        0.06499999761581421,
        0.0637499988079071,
        0.0625,
        0.04625000059604645,
        0.07750000059604645
      ],
      "collision_mean": 0.05828124936670065,
      "mean_gap": 0.9999973326921463,
      "cfg": {
        "lr": 0.003,
        "weight_decay": 0.0001
      }
    }
  },
  "idea": {
    "mean": 0.08495018165558577,
    "std": 0.0036923579732418346,
    "per_seed": [
      0.08852718770503998,
      0.08314542472362518,
      0.08323974907398224,
      0.0792478397488594,
      0.0855635404586792,
      0.09225999563932419,
      0.08457005023956299,
      0.08304766565561295
    ],
    "n": 8,
    "collision_per_seed": [
      0.07499999552965164,
      0.05624999850988388,
      0.05624999850988388,
      0.07374999672174454,
      0.05999999865889549,
      0.07000000029802322,
      0.0625,
      0.06875000149011612
    ],
    "collision_mean": 0.06531249871477485,
    "mean_gap": 0.9999975040555,
    "cfg": {
      "lr": 0.003,
      "weight_decay": 0.0001,
      "lam": 0.1,
      "beta": 2.0,
      "delta": 0.08
    }
  },
  "comparison": {
    "delta_mean": 0.013171124272048473,
    "idea_wins": 0,
    "n_pairs": 8,
    "per_seed_diffs": [
      0.013828955590724945,
      0.014328934252262115,
      0.014567740261554718,
      0.013787157833576202,
      0.014008529484272003,
      0.013736352324485779,
      0.012276388704776764,
      0.00883493572473526
    ],
    "p_value": 0.0081,
    "verdict": "idea worse (significant)",
    "system_worked": false
  },
  "custom_track": {
    "name": "unordered_pointset_denoising",
    "file": "pointset_fusion_track.py",
    "domain": "point-set-diffusion"
  },
  "idea_settings": [
    {
      "mean": 0.08495018165558577,
      "std": 0.0036923579732418346,
      "per_seed": [
        0.08852718770503998,
        0.08314542472362518,
        0.08323974907398224,
        0.0792478397488594,
        0.0855635404586792,
        0.09225999563932419,
        0.08457005023956299,
        0.08304766565561295
      ],
      "n": 8,
      "collision_per_seed": [
        0.07499999552965164,
        0.05624999850988388,
        0.05624999850988388,
        0.07374999672174454,
        0.05999999865889549,
        0.07000000029802322,
        0.0625,
        0.06875000149011612
      ],
      "collision_mean": 0.06531249871477485,
      "mean_gap": 0.9999975040555,
      "cfg": {
        "lr": 0.003,
        "weight_decay": 0.0001,
        "lam": 0.1,
        "beta": 2.0,
        "delta": 0.08
      }
    },
    {
      "mean": 0.0936331320554018,
      "std": 0.0037630771536154607,
      "per_seed": [
        0.09448400139808655,
        0.09308362752199173,
        0.08933884650468826,
        0.0864451453089714,
        0.09648385643959045,
        0.09782309830188751,
        0.09371592104434967,
        0.09769055992364883
      ],
      "n": 8,
      "collision_per_seed": [
        0.07999999821186066,
        0.07000000029802322,
        0.06749999523162842,
        0.07124999910593033,
        0.0612499974668026,
        0.07249999791383743,
        0.0637499988079071,
        0.06499999761581421
      ],
      "collision_mean": 0.0689062480814755,
      "mean_gap": 0.9999975264072418,
      "cfg": {
        "lr": 0.002,
        "weight_decay": 0.0001,
        "lam": 0.1,
        "beta": 2.0,
        "delta": 0.08
      }
    },
    {
      "mean": 0.09946490358561277,
      "std": 0.0037473580173809696,
      "per_seed": [
        0.09996496886014938,
        0.10018042474985123,
        0.09583928436040878,
        0.09132290631532669,
        0.1030997484922409,
        0.10119754076004028,
        0.1009417474269867,
        0.10317260771989822
      ],
      "n": 8,
      "collision_per_seed": [
        0.07750000059604645,
        0.08374999463558197,
        0.05874999985098839,
        0.06499999761581421,
        0.06749999523162842,
        0.07249999791383743,
        0.08249999582767487,
        0.07625000178813934
      ],
      "collision_mean": 0.07296874793246388,
      "mean_gap": 0.9999975264072418,
      "cfg": {
        "lr": 0.001,
        "weight_decay": 0.0001,
        "lam": 0.1,
        "beta": 2.0,
        "delta": 0.08
      }
    }
  ],
  "mechanism_signature": {
    "quantity": "normalized predicted short-gap collision rate",
    "baseline_predicted_collision": 0.05828124936670065,
    "idea_predicted_collision": 0.06531249871477485,
    "predicted_delta": 0.007031249348074198,
    "observed_test_mse_delta": 0.013171124272048473,
    "confirmed": false
  }
}