1{
  2  "bench_version": 1,
  3  "track": "sequence",
  4  "model": "transformer_tiny",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.001
 10    },
 11    "sweep": [
 12      {
 13        "cfg": {
 14          "lr": 0.001
 15        },
 16        "mean": 0.19863465800881386
 17      },
 18      {
 19        "cfg": {
 20          "lr": 0.003
 21        },
 22        "mean": 0.20291408896446228
 23      },
 24      {
 25        "cfg": {
 26          "lr": 0.01
 27        },
 28        "mean": 0.7486049979925156
 29      }
 30    ],
 31    "full": {
 32      "mean": 0.20585612207651138,
 33      "std": 0.022614501649959134,
 34      "per_seed": [
 35        0.23615866899490356,
 36        0.18743106722831726,
 37        0.17751933634281158,
 38        0.19342955946922302,
 39        0.20024268329143524,
 40        0.20917430520057678,
 41        0.24740388989448547,
 42        0.19548946619033813
 43      ],
 44      "n": 8
 45    }
 46  },
 47  "idea": {
 48    "mean": 0.14687445014715195,
 49    "std": 0.018827814048492597,
 50    "per_seed": [
 51      0.17195841670036316,
 52      0.1316526234149933,
 53      0.12245547771453857,
 54      0.15120451152324677,
 55      0.135431170463562,
 56      0.12997840344905853,
 57      0.176153764128685,
 58      0.15616123378276825
 59    ],
 60    "n": 8
 61  },
 62  "comparison": {
 63    "delta_mean": -0.058981671929359436,
 64    "idea_wins": 8,
 65    "n_pairs": 8,
 66    "per_seed_diffs": [
 67      -0.0642002522945404,
 68      -0.055778443813323975,
 69      -0.05506385862827301,
 70      -0.04222504794597626,
 71      -0.06481151282787323,
 72      -0.07919590175151825,
 73      -0.07125012576580048,
 74      -0.039328232407569885
 75    ],
 76    "p_value": 0.0081,
 77    "mde": 0.011451699601831964,
 78    "mde_rel_pct": 5.5629628530433815,
 79    "verdict": "idea better (significant)",
 80    "system_worked": true
 81  },
 82  "idea_sweep": [
 83    {
 84      "cfg": {
 85        "lr": 0.001
 86      },
 87      "full": {
 88        "mean": 0.14746149070560932,
 89        "std": 0.012264365784412576,
 90        "per_seed": [
 91          0.16718325018882751,
 92          0.12648990750312805,
 93          0.13353104889392853,
 94          0.15779778361320496,
 95          0.14990615844726562,
 96          0.14269356429576874,
 97          0.15414929389953613,
 98          0.14794091880321503
 99        ],
100        "n": 8
101      }
102    },
103    {
104      "cfg": {
105        "lr": 0.003
106      },
107      "full": {
108        "mean": 0.14687445014715195,
109        "std": 0.018827814048492597,
110        "per_seed": [
111          0.17195841670036316,
112          0.1316526234149933,
113          0.12245547771453857,
114          0.15120451152324677,
115          0.135431170463562,
116          0.12997840344905853,
117          0.176153764128685,
118          0.15616123378276825
119        ],
120        "n": 8
121      }
122    },
123    {
124      "cfg": {
125        "lr": 0.01
126      },
127      "full": {
128        "mean": 0.23110334761440754,
129        "std": 0.04669852475794671,
130        "per_seed": [
131          0.23717589676380157,
132          0.19244495034217834,
133          0.21007037162780762,
134          0.2582976818084717,
135          0.203369140625,
136          0.19808459281921387,
137          0.3420470058917999,
138          0.2073371410369873
139        ],
140        "n": 8
141      }
142    }
143  ],
144  "shared_architecture": true,
145  "gain_graph_edges": 51,
146  "mechanism_signature": {
147    "prediction": "trained gain-rigid sparse attention should retain connected input influence, with nonzero gradient reach comparable to dense attention",
148    "observed": {
149      "seed": 0,
150      "baseline_test_mse": 0.22468174993991852,
151      "idea_test_mse": 0.17195841670036316,
152      "baseline_input_gradient_mean": 0.1916990876197815,
153      "idea_input_gradient_mean": 0.27051663398742676,
154      "baseline_input_gradient_active_fraction": 1.0,
155      "idea_input_gradient_active_fraction": 1.0,
156      "gain_edges": 51,
157      "nodes": 32
158    },
159    "confirmed": true
160  }
161}