{
  "bench_version": 1,
  "track": "tabular",
  "model": "mlp_tiny",
  "metric_direction": "lower is better",
  "n_seeds": 8,
  "baseline": {
    "best_cfg": {
      "lr": 0.01,
      "weight_decay": 0.0001
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.001,
          "weight_decay": 0.0
        },
        "mean": 15.90550446510315
      },
      {
        "cfg": {
          "lr": 0.001,
          "weight_decay": 0.0001
        },
        "mean": 15.877723932266235
      },
      {
        "cfg": {
          "lr": 0.003,
          "weight_decay": 0.0
        },
        "mean": 9.71069884300232
      },
      {
        "cfg": {
          "lr": 0.003,
          "weight_decay": 0.0001
        },
        "mean": 9.578425407409668
      },
      {
        "cfg": {
          "lr": 0.01,
          "weight_decay": 0.0
        },
        "mean": 7.003860950469971
      },
      {
        "cfg": {
          "lr": 0.01,
          "weight_decay": 0.0001
        },
        "mean": 6.940991640090942
      }
    ],
    "full": {
      "mean": 7.183188855648041,
      "std": 0.6070482381931488,
      "per_seed": [
        6.755654335021973,
        7.163381099700928,
        6.924261093139648,
        6.920670032501221,
        6.168270111083984,
        7.485415458679199,
        7.807327747344971,
        8.240530967712402
      ],
      "n": 8
    }
  },
  "idea": {
    "mean": 220.65338706970215,
    "std": 6.711567657993523,
    "per_seed": [
      227.8157958984375,
      220.93875122070312,
      227.989990234375,
      210.31065368652344,
      223.4892578125,
      218.9625701904297,
      210.06019592285156,
      225.65988159179688
    ],
    "n": 8
  },
  "comparison": {
    "delta_mean": 213.4701982140541,
    "idea_wins": 0,
    "n_pairs": 8,
    "per_seed_diffs": [
      221.06014156341553,
      213.7753701210022,
      221.06572914123535,
      203.38998365402222,
      217.32098770141602,
      211.4771547317505,
      202.2528681755066,
      217.41935062408447
    ],
    "p_value": 0.0081,
    "mde": 6.138174864466896,
    "mde_rel_pct": 85.45194881853251,
    "verdict": "idea worse (significant)",
    "system_worked": false
  },
  "mechanism_signature": {
    "prediction": "noise enters acceleration only; at temperature=0 injected-noise std is zero",
    "predicted_injected_noise_std": 0.0,
    "observed_trained_acceleration_norm_mean": 15.645100752559667,
    "observed_trained_gradient_norm_mean": 98.2094500956436,
    "observed_acceleration_to_gradient_ratio_mean": 0.1593389100379601,
    "confirmed": false,
    "note": "State acceleration is nonzero from deterministic gradients, so this does not confirm a zero total-state prediction; it directly measures trained-model behavior."
  },
  "idea_sweep": [
    {
      "cfg": {
        "lr": 0.001,
        "gamma": 0.5,
        "temperature": 0.0
      },
      "result": {
        "mean": 235.77590942382812,
        "std": 7.662576847576715,
        "per_seed": [
          243.6631317138672,
          235.733154296875,
          241.19017028808594,
          221.46832275390625,
          245.36273193359375,
          231.23699951171875,
          228.62933349609375,
          238.92343139648438
        ],
        "n": 8
      }
    },
    {
      "cfg": {
        "lr": 0.003,
        "gamma": 1.0,
        "temperature": 0.0
      },
      "result": {
        "mean": 220.65338706970215,
        "std": 6.711567657993523,
        "per_seed": [
          227.8157958984375,
          220.93875122070312,
          227.989990234375,
          210.31065368652344,
          223.4892578125,
          218.9625701904297,
          210.06019592285156,
          225.65988159179688
        ],
        "n": 8
      }
    },
    {
      "cfg": {
        "lr": 0.01,
        "gamma": 2.0,
        "temperature": 0.0
      },
      "result": {
        "mean": 471.48626041412354,
        "std": 384.5950909075575,
        "per_seed": [
          105.4653091430664,
          324.9513854980469,
          402.4331970214844,
          1098.881591796875,
          90.00657653808594,
          657.0219116210938,
          78.45921325683594,
          1014.6708984375
        ],
        "n": 8
      }
    }
  ],
  "protocol_notes": {
    "paired_seeds": [
      0,
      1,
      2,
      3,
      4,
      5,
      6,
      7
    ],
    "baseline_sweep_seeds": [
      0,
      1,
      2,
      3
    ],
    "dataset_sizes": [
      400,
      400
    ],
    "epochs": 30,
    "batch": 128,
    "baseline": "Adam via bench.train_model",
    "idea": "third-order Langevin; only optimizer loop differs",
    "search_space_union_learning_rates": [
      0.001,
      0.003,
      0.01
    ]
  }
}