1{
  2  "bench_version": 1,
  3  "track": "poisson_boundary",
  4  "model": "scalar_potential_backbone",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.03
 10    },
 11    "sweep": [
 12      {
 13        "cfg": {
 14          "lr": 0.003
 15        },
 16        "mean": 0.12233347818255424
 17      },
 18      {
 19        "cfg": {
 20          "lr": 0.01
 21        },
 22        "mean": 0.05105730285868049
 23      },
 24      {
 25        "cfg": {
 26          "lr": 0.03
 27        },
 28        "mean": 0.009451728663407266
 29      }
 30    ],
 31    "full": {
 32      "mean": 0.01093649520771578,
 33      "std": 0.003025037049171694,
 34      "per_seed": [
 35        0.00799343828111887,
 36        0.006817351561039686,
 37        0.010449049063026905,
 38        0.012547075748443604,
 39        0.008240480907261372,
 40        0.011241904459893703,
 41        0.01626167818903923,
 42        0.013940983451902866
 43      ],
 44      "n": 8
 45    }
 46  },
 47  "idea": {
 48    "mean": 0.03724087378941476,
 49    "std": 0.012178382868521676,
 50    "per_seed": [
 51      0.030537452548742294,
 52      0.06148594617843628,
 53      0.022630251944065094,
 54      0.05179433524608612,
 55      0.032382115721702576,
 56      0.02730206958949566,
 57      0.03700961917638779,
 58      0.0347851999104023
 59    ],
 60    "n": 8
 61  },
 62  "comparison": {
 63    "delta_mean": 0.026304378581698984,
 64    "idea_wins": 0,
 65    "n_pairs": 8,
 66    "per_seed_diffs": [
 67      0.022544014267623425,
 68      0.05466859461739659,
 69      0.012181202881038189,
 70      0.03924725949764252,
 71      0.024141634814441204,
 72      0.016060165129601955,
 73      0.020747940987348557,
 74      0.02084421645849943
 75    ],
 76    "p_value": 0.0081,
 77    "mde": 0.01163752387239483,
 78    "mde_rel_pct": 106.40999379933409,
 79    "verdict": "idea worse (significant)",
 80    "system_worked": false
 81  },
 82  "mechanism_signature": {
 83    "stage1_prediction": "d squared equals zero for the learned potential gradient",
 84    "baseline_gradient_curl_rms": 1.36577753551137e-07,
 85    "predicted_gradient_curl_rms": 0.0,
 86    "observed_gradient_curl_rms": 1.154990201257533e-07,
 87    "n_probe": 128,
 88    "confirmed": true,
 89    "idea_sweep": [
 90      {
 91        "cfg": {
 92          "lr": 0.003
 93        },
 94        "result": {
 95          "mean": 0.11393675301223993,
 96          "std": 0.01067011604185375,
 97          "per_seed": [
 98            0.11674976348876953,
 99            0.11242678761482239,
100            0.09828850626945496,
101            0.12337931990623474,
102            0.13426120579242706,
103            0.11549603939056396,
104            0.10730288922786713,
105            0.1035895124077797
106          ],
107          "n": 8
108        }
109      },
110      {
111        "cfg": {
112          "lr": 0.01
113        },
114        "result": {
115          "mean": 0.08499461971223354,
116          "std": 0.00870385202148945,
117          "per_seed": [
118            0.07853147387504578,
119            0.0928913876414299,
120            0.07291875034570694,
121            0.09869463741779327,
122            0.09521957486867905,
123            0.08183202892541885,
124            0.08108515292406082,
125            0.07878395169973373
126          ],
127          "n": 8
128        }
129      },
130      {
131        "cfg": {
132          "lr": 0.03
133        },
134        "result": {
135          "mean": 0.03724087378941476,
136          "std": 0.012178382868521676,
137          "per_seed": [
138            0.030537452548742294,
139            0.06148594617843628,
140            0.022630251944065094,
141            0.05179433524608612,
142            0.032382115721702576,
143            0.02730206958949566,
144            0.03700961917638779,
145            0.0347851999104023
146          ],
147          "n": 8
148        }
149      }
150    ],
151    "math_check": {
152      "nilpotence_max_abs": 0.0,
153      "derivative_identity_max_abs": 0.0
154    }
155  },
156  "math_check": {
157    "nilpotence_max_abs": 0.0,
158    "derivative_identity_max_abs": 0.0
159  },
160  "idea_selected_cfg": {
161    "lr": 0.03
162  }
163}