1{
  2  "bench_version": 1,
  3  "track": "soft_graph_connectivity",
  4  "model": "graphnet_custom",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "alpha": 0.0,
 10      "lr": 0.001
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "alpha": 0.0,
 16          "lr": 0.001
 17        },
 18        "mean": 1.6224595356106875e-05
 19      },
 20      {
 21        "cfg": {
 22          "alpha": 0.0,
 23          "lr": 0.003
 24        },
 25        "mean": 1.6673563322910923e-05
 26      },
 27      {
 28        "cfg": {
 29          "alpha": 0.0,
 30          "lr": 0.01
 31        },
 32        "mean": 2.209703552580322e-05
 33      },
 34      {
 35        "cfg": {
 36          "alpha": 0.0,
 37          "lr": 0.002
 38        },
 39        "mean": 1.656733707022795e-05
 40      },
 41      {
 42        "cfg": {
 43          "alpha": 0.0,
 44          "lr": 0.006
 45        },
 46        "mean": 1.8259379430674016e-05
 47      }
 48    ],
 49    "full": {
 50      "mean": 1.7648152720539656e-05,
 51      "std": 2.347719277139452e-06,
 52      "per_seed": [
 53        1.82739404408494e-05,
 54        1.5822064597159624e-05,
 55        1.3746795048064087e-05,
 56        1.7055581338354386e-05,
 57        1.827272899390664e-05,
 58        1.6145058907568455e-05,
 59        2.0411813238752075e-05,
 60        2.145723919966258e-05
 61      ],
 62      "n": 8
 63    },
 64    "idea_side_sweep": [
 65      {
 66        "cfg": {
 67          "alpha": 0.0001,
 68          "lr": 0.001
 69        },
 70        "mean": 1.811643392102269e-05
 71      },
 72      {
 73        "cfg": {
 74          "alpha": 0.001,
 75          "lr": 0.001
 76        },
 77        "mean": 0.00017178840425913222
 78      },
 79      {
 80        "cfg": {
 81          "alpha": 0.01,
 82          "lr": 0.001
 83        },
 84        "mean": 0.009164006682112813
 85      },
 86      {
 87        "cfg": {
 88          "alpha": 0.0001,
 89          "lr": 0.002
 90        },
 91        "mean": 1.70701698607445e-05
 92      },
 93      {
 94        "cfg": {
 95          "alpha": 0.001,
 96          "lr": 0.002
 97        },
 98        "mean": 0.00016471587878186256
 99      },
100      {
101        "cfg": {
102          "alpha": 0.01,
103          "lr": 0.002
104        },
105        "mean": 0.009119202382862568
106      },
107      {
108        "cfg": {
109          "alpha": 0.0001,
110          "lr": 0.006
111        },
112        "mean": 1.849937916631461e-05
113      },
114      {
115        "cfg": {
116          "alpha": 0.001,
117          "lr": 0.006
118        },
119        "mean": 0.00016692252029315569
120      },
121      {
122        "cfg": {
123          "alpha": 0.01,
124          "lr": 0.006
125        },
126        "mean": 0.00913658645004034
127      }
128    ]
129  },
130  "idea": {
131    "mean": 1.8547835793469858e-05,
132    "std": 3.1568402610061954e-06,
133    "per_seed": [
134      1.927832818182651e-05,
135      1.6839227100717835e-05,
136      1.4905762327543925e-05,
137      1.7257361832889728e-05,
138      1.818846612877678e-05,
139      1.4989042028901167e-05,
140      2.3812290237401612e-05,
141      2.3112208509701304e-05
142    ],
143    "n": 8
144  },
145  "comparison": {
146    "delta_mean": 8.996830729302019e-07,
147    "idea_wins": 2,
148    "n_pairs": 8,
149    "per_seed_diffs": [
150      1.0043877409771085e-06,
151      1.017162503558211e-06,
152      1.158967279479839e-06,
153      2.0178049453534186e-07,
154      -8.426286512985826e-08,
155      -1.1560168786672875e-06,
156      3.4004769986495376e-06,
157      1.654969310038723e-06
158    ],
159    "p_value": 0.1026,
160    "mde": 1.1217084950015763e-06,
161    "mde_rel_pct": 6.355954148651972,
162    "verdict": "no significant win",
163    "system_worked": false
164  },
165  "custom_track": {
166    "name": "soft_graph_connectivity",
167    "file": "graph_track.py",
168    "domain": "graph-nn"
169  },
170  "math_check": {
171    "relative_rank_one_pdet_error": 3.280937793951094e-16,
172    "relative_matrix_tree_cofactor_error": 1.6404688969755466e-16
173  },
174  "idea_cfg": {
175    "alpha": 0.0001,
176    "lr": 0.002
177  },
178  "mechanism_signature": {
179    "quantity": "Fiedler eigenvalue of trained predicted adjacency",
180    "baseline_predicted_fiedler_mean": 1.6474629960954188,
181    "idea_predicted_fiedler_mean": 1.6093307891239723,
182    "baseline_components_mean": 1.0,
183    "idea_components_mean": 1.0,
184    "prediction": "tree regularization should increase Fiedler connectivity",
185    "confirmed": false
186  }
187}