{
  "bench_version": 1,
  "track": "sequence",
  "model": "transformer_tiny",
  "metric_direction": "lower is better",
  "n_seeds": 8,
  "baseline": {
    "best_cfg": {
      "lr": 0.001
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.001
        },
        "mean": 0.19863465800881386
      },
      {
        "cfg": {
          "lr": 0.003
        },
        "mean": 0.20291408896446228
      },
      {
        "cfg": {
          "lr": 0.01
        },
        "mean": 0.7486049979925156
      }
    ],
    "full": {
      "mean": 0.20585612207651138,
      "std": 0.022614501649959134,
      "per_seed": [
        0.23615866899490356,
        0.18743106722831726,
        0.17751933634281158,
        0.19342955946922302,
        0.20024268329143524,
        0.20917430520057678,
        0.24740388989448547,
        0.19548946619033813
      ],
      "n": 8
    }
  },
  "idea": {
    "mean": 0.14687445014715195,
    "std": 0.018827814048492597,
    "per_seed": [
      0.17195841670036316,
      0.1316526234149933,
      0.12245547771453857,
      0.15120451152324677,
      0.135431170463562,
      0.12997840344905853,
      0.176153764128685,
      0.15616123378276825
    ],
    "n": 8
  },
  "comparison": {
    "delta_mean": -0.058981671929359436,
    "idea_wins": 8,
    "n_pairs": 8,
    "per_seed_diffs": [
      -0.0642002522945404,
      -0.055778443813323975,
      -0.05506385862827301,
      -0.04222504794597626,
      -0.06481151282787323,
      -0.07919590175151825,
      -0.07125012576580048,
      -0.039328232407569885
    ],
    "p_value": 0.0081,
    "mde": 0.011451699601831964,
    "mde_rel_pct": 5.5629628530433815,
    "verdict": "idea better (significant)",
    "system_worked": true
  },
  "idea_sweep": [
    {
      "cfg": {
        "lr": 0.001
      },
      "full": {
        "mean": 0.14746149070560932,
        "std": 0.012264365784412576,
        "per_seed": [
          0.16718325018882751,
          0.12648990750312805,
          0.13353104889392853,
          0.15779778361320496,
          0.14990615844726562,
          0.14269356429576874,
          0.15414929389953613,
          0.14794091880321503
        ],
        "n": 8
      }
    },
    {
      "cfg": {
        "lr": 0.003
      },
      "full": {
        "mean": 0.14687445014715195,
        "std": 0.018827814048492597,
        "per_seed": [
          0.17195841670036316,
          0.1316526234149933,
          0.12245547771453857,
          0.15120451152324677,
          0.135431170463562,
          0.12997840344905853,
          0.176153764128685,
          0.15616123378276825
        ],
        "n": 8
      }
    },
    {
      "cfg": {
        "lr": 0.01
      },
      "full": {
        "mean": 0.23110334761440754,
        "std": 0.04669852475794671,
        "per_seed": [
          0.23717589676380157,
          0.19244495034217834,
          0.21007037162780762,
          0.2582976818084717,
          0.203369140625,
          0.19808459281921387,
          0.3420470058917999,
          0.2073371410369873
        ],
        "n": 8
      }
    }
  ],
  "shared_architecture": true,
  "gain_graph_edges": 51,
  "mechanism_signature": {
    "prediction": "trained gain-rigid sparse attention should retain connected input influence, with nonzero gradient reach comparable to dense attention",
    "observed": {
      "seed": 0,
      "baseline_test_mse": 0.22468174993991852,
      "idea_test_mse": 0.17195841670036316,
      "baseline_input_gradient_mean": 0.1916990876197815,
      "idea_input_gradient_mean": 0.27051663398742676,
      "baseline_input_gradient_active_fraction": 1.0,
      "idea_input_gradient_active_fraction": 1.0,
      "gain_edges": 51,
      "nodes": 32
    },
    "confirmed": true
  }
}