1{
  2  "bench_version": 1,
  3  "track": "geometric_graph_diffusion",
  4  "model": "GeometricMessageNet",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.003,
 10      "alpha": 0.5
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "alpha": 0.5
 17        },
 18        "mean": 0.08639695774763823
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "alpha": 0.5
 24        },
 25        "mean": 0.06887020450085402
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.01,
 30          "alpha": 0.5
 31        },
 32        "mean": 0.08051598258316517
 33      }
 34    ],
 35    "full": {
 36      "mean": 0.08486439986154437,
 37      "std": 0.029581161321461158,
 38      "per_seed": [
 39        0.036324214190244675,
 40        0.07466080784797668,
 41        0.09419873356819153,
 42        0.07029706239700317,
 43        0.14098234474658966,
 44        0.07441659271717072,
 45        0.0738333985209465,
 46        0.11420204490423203
 47      ],
 48      "n": 8
 49    }
 50  },
 51  "idea": {
 52    "mean": 0.07120458083227277,
 53    "std": 0.023789005397093055,
 54    "per_seed": [
 55      0.048530153930187225,
 56      0.06982574611902237,
 57      0.07111480087041855,
 58      0.05362492427229881,
 59      0.1285717934370041,
 60      0.08061950653791428,
 61      0.05659816414117813,
 62      0.06075155735015869
 63    ],
 64    "n": 8
 65  },
 66  "comparison": {
 67    "delta_mean": -0.013659819029271603,
 68    "idea_wins": 6,
 69    "n_pairs": 8,
 70    "per_seed_diffs": [
 71      0.01220593973994255,
 72      -0.004835061728954315,
 73      -0.02308393269777298,
 74      -0.01667213812470436,
 75      -0.012410551309585571,
 76      0.006202913820743561,
 77      -0.01723523437976837,
 78      -0.053450487554073334
 79    ],
 80    "p_value": 0.0868,
 81    "mde": 0.016833120021636148,
 82    "mde_rel_pct": 19.835313805434616,
 83    "verdict": "no significant win",
 84    "system_worked": false
 85  },
 86  "mechanism_signature": {
 87    "track_structure": "geometric graph node-field diffusion regression",
 88    "trained_model_measurements": {
 89      "baseline": {
 90        "test_mse": 0.036324214190244675,
 91        "predicted_l2": 0.22030678391456604,
 92        "observed_gradient_l2": 0.26773717999458313,
 93        "predicted_observed_corr": 0.8789346780854339
 94      },
 95      "idea": {
 96        "test_mse": 0.039621297270059586,
 97        "predicted_l2": 0.2222180962562561,
 98        "observed_gradient_l2": 0.24487970769405365,
 99        "predicted_observed_corr": 0.38409918113832164
100      },
101      "prediction": "distance weighting preserves more long-range influence",
102      "confirmed": false
103    },
104    "idea_sweep": [
105      {
106        "cfg": {
107          "lr": 0.001,
108          "alpha": 0.5
109        },
110        "result": {
111          "mean": 0.07120458083227277,
112          "std": 0.023789005397093055,
113          "per_seed": [
114            0.048530153930187225,
115            0.06982574611902237,
116            0.07111480087041855,
117            0.05362492427229881,
118            0.1285717934370041,
119            0.08061950653791428,
120            0.05659816414117813,
121            0.06075155735015869
122          ],
123          "n": 8
124        }
125      },
126      {
127        "cfg": {
128          "lr": 0.003,
129          "alpha": 0.5
130        },
131        "result": {
132          "mean": 0.07151469262316823,
133          "std": 0.05479380310064357,
134          "per_seed": [
135            0.039621297270059586,
136            0.06278985738754272,
137            0.053295787423849106,
138            0.06647685915231705,
139            0.21443338692188263,
140            0.05103813484311104,
141            0.044252004474401474,
142            0.040210213512182236
143          ],
144          "n": 8
145        }
146      },
147      {
148        "cfg": {
149          "lr": 0.01,
150          "alpha": 0.5
151        },
152        "result": {
153          "mean": 0.07402465306222439,
154          "std": 0.07012354046889444,
155          "per_seed": [
156            0.03252619504928589,
157            0.0537540428340435,
158            0.06584807485342026,
159            0.050242867320775986,
160            0.2579071521759033,
161            0.04949888214468956,
162            0.042814504355192184,
163            0.039605505764484406
164          ],
165          "n": 8
166        }
167      }
168    ],
169    "confirmed": false
170  },
171  "custom_track": {
172    "name": "geometric_graph_diffusion",
173    "file": "graph_track.py",
174    "domain": "graph-neural-networks"
175  },
176  "protocol_notes": {
177    "paired_seeds": [
178      0,
179      1,
180      2,
181      3,
182      4,
183      5,
184      6,
185      7
186    ],
187    "baseline_and_idea_same_architecture": true,
188    "only_intervention": "edge weight d^2 versus unit edge weight",
189    "device": "cuda"
190  }
191}