{
  "bench_version": 1,
  "track": "conditional_sequence_diffusion",
  "model": "score_mlp",
  "metric_direction": "lower is better",
  "n_seeds": 8,
  "baseline": {
    "best_cfg": {
      "lr": 0.003,
      "epochs": 12,
      "steps": 16
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.001,
          "epochs": 12,
          "steps": 8
        },
        "mean": 0.3794306069612503
      },
      {
        "cfg": {
          "lr": 0.001,
          "epochs": 12,
          "steps": 16
        },
        "mean": 0.36941855028271675
      },
      {
        "cfg": {
          "lr": 0.003,
          "epochs": 12,
          "steps": 8
        },
        "mean": 0.3344487212598324
      },
      {
        "cfg": {
          "lr": 0.003,
          "epochs": 12,
          "steps": 16
        },
        "mean": 0.32668623328208923
      },
      {
        "cfg": {
          "lr": 0.01,
          "epochs": 12,
          "steps": 8
        },
        "mean": 0.3527134172618389
      },
      {
        "cfg": {
          "lr": 0.01,
          "epochs": 12,
          "steps": 16
        },
        "mean": 0.3493003807961941
      }
    ],
    "full": {
      "mean": 0.32668623328208923,
      "std": 0.0183314657731986,
      "per_seed": [
        0.34015029668807983,
        0.3335983455181122,
        0.34249189496040344,
        0.2935677170753479,
        0.3214184641838074,
        0.3177299499511719,
        0.35397353768348694,
        0.3105596601963043
      ],
      "n": 8
    }
  },
  "idea": {
    "mean": 0.3268098309636116,
    "std": 0.018441350747342378,
    "per_seed": [
      0.3401668965816498,
      0.33357036113739014,
      0.3425343632698059,
      0.2938295602798462,
      0.32191580533981323,
      0.3179206848144531,
      0.3546198308467865,
      0.30992114543914795
    ],
    "n": 8,
    "trials": [
      {
        "cfg": {
          "lr": 0.001,
          "epochs": 12,
          "steps": 8
        },
        "mean": 0.3800293132662773
      },
      {
        "cfg": {
          "lr": 0.001,
          "epochs": 12,
          "steps": 16
        },
        "mean": 0.3697975389659405
      },
      {
        "cfg": {
          "lr": 0.003,
          "epochs": 12,
          "steps": 8
        },
        "mean": 0.33464548736810684
      },
      {
        "cfg": {
          "lr": 0.003,
          "epochs": 12,
          "steps": 16
        },
        "mean": 0.3268098309636116
      },
      {
        "cfg": {
          "lr": 0.01,
          "epochs": 12,
          "steps": 8
        },
        "mean": 0.35246242582798004
      },
      {
        "cfg": {
          "lr": 0.01,
          "epochs": 12,
          "steps": 16
        },
        "mean": 0.34918101504445076
      }
    ]
  },
  "comparison": {
    "delta_mean": 0.00012359768152236938,
    "idea_wins": 2,
    "n_pairs": 8,
    "per_seed_diffs": [
      1.659989356994629e-05,
      -2.79843807220459e-05,
      4.246830940246582e-05,
      0.000261843204498291,
      0.0004973411560058594,
      0.00019073486328125,
      0.0006462931632995605,
      -0.0006385147571563721
    ],
    "p_value": 0.34835,
    "mde": 0.00032512406414669,
    "mde_rel_pct": 0.09952181360086565,
    "verdict": "no measurable effect",
    "system_worked": false,
    "permutation_pvalue": 0.34835
  },
  "mechanism_signature": {
    "density_mean": 0.005077726362893979,
    "density_max": 0.005121961701661348,
    "width_inverse_sqrt_corr": 0.9998106126499892,
    "observed_cost_cv": 7.354513075912003e-05,
    "predicted_cost_equalization": "low CV under sqrt allocation",
    "confirmed": true
  },
  "custom_track": {
    "name": "conditional_sequence_diffusion",
    "file": "custom_diffusion_track.py",
    "domain": "sequence-level diffusion sampling"
  }
}