1{
  2  "bench_version": 1,
  3  "track": "tangent_surface_vector_regression",
  4  "model": "shared_mlp",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.01
 10    },
 11    "sweep": [
 12      {
 13        "cfg": {
 14          "lr": 0.001
 15        },
 16        "mean": 0.03764702659100294
 17      },
 18      {
 19        "cfg": {
 20          "lr": 0.003
 21        },
 22        "mean": 0.030783741734921932
 23      },
 24      {
 25        "cfg": {
 26          "lr": 0.01
 27        },
 28        "mean": 0.026880070567131042
 29      }
 30    ],
 31    "union_full": [
 32      {
 33        "cfg": {
 34          "lr": 0.001
 35        },
 36        "full": {
 37          "mean": 0.03800344513729215,
 38          "std": 0.0036767906806796174,
 39          "per_seed": [
 40            0.03472031280398369,
 41            0.03560163080692291,
 42            0.042614057660102844,
 43            0.03765210509300232,
 44            0.04322683811187744,
 45            0.03853558748960495,
 46            0.03991776704788208,
 47            0.03175926208496094
 48          ],
 49          "n": 8
 50        }
 51      },
 52      {
 53        "cfg": {
 54          "lr": 0.003
 55        },
 56        "full": {
 57          "mean": 0.03143041324801743,
 58          "std": 0.002817096624089596,
 59          "per_seed": [
 60            0.029847173020243645,
 61            0.032056987285614014,
 62            0.0286399032920599,
 63            0.03259090334177017,
 64            0.038180507719516754,
 65            0.030082618817687035,
 66            0.03025837428867817,
 67            0.029786838218569756
 68          ],
 69          "n": 8
 70        }
 71      },
 72      {
 73        "cfg": {
 74          "lr": 0.01
 75        },
 76        "full": {
 77          "mean": 0.02747795032337308,
 78          "std": 0.0043431611845433085,
 79          "per_seed": [
 80            0.02603250928223133,
 81            0.02745499461889267,
 82            0.026675675064325333,
 83            0.027357103303074837,
 84            0.038280390202999115,
 85            0.02431103214621544,
 86            0.0268161091953516,
 87            0.02289578877389431
 88          ],
 89          "n": 8
 90        }
 91      }
 92    ],
 93    "full": {
 94      "mean": 0.02747795032337308,
 95      "std": 0.0043431611845433085,
 96      "per_seed": [
 97        0.02603250928223133,
 98        0.02745499461889267,
 99        0.026675675064325333,
100        0.027357103303074837,
101        0.038280390202999115,
102        0.02431103214621544,
103        0.0268161091953516,
104        0.02289578877389431
105      ],
106      "n": 8
107    }
108  },
109  "idea": {
110    "mean": 0.021506645949557424,
111    "std": 0.004870355712475741,
112    "per_seed": [
113      0.014883503317832947,
114      0.020688023418188095,
115      0.030457986518740654,
116      0.023641860112547874,
117      0.021189182996749878,
118      0.022041697055101395,
119      0.024708442389965057,
120      0.014442471787333488
121    ],
122    "n": 8
123  },
124  "comparison": {
125    "delta_mean": -0.005971304373815656,
126    "idea_wins": 7,
127    "n_pairs": 8,
128    "per_seed_diffs": [
129      -0.011149005964398384,
130      -0.006766971200704575,
131      0.0037823114544153214,
132      -0.0037152431905269623,
133      -0.017091207206249237,
134      -0.002269335091114044,
135      -0.0021076668053865433,
136      -0.008453316986560822
137    ],
138    "p_value": 0.03805,
139    "mde": 0.005345683018986985,
140    "mde_rel_pct": 19.45444604155894,
141    "verdict": "idea better (significant)",
142    "system_worked": true
143  },
144  "custom_track": {
145    "name": "tangent_surface_vector_regression",
146    "file": "tangent_surface_track.py",
147    "domain": "geometry/point-cloud"
148  },
149  "idea_sweep": [
150    {
151      "cfg": {
152        "lr": 0.001
153      },
154      "full": {
155        "mean": 0.02933011995628476,
156        "std": 0.005050140438499371,
157        "per_seed": [
158          0.028748154640197754,
159          0.032053809612989426,
160          0.03960653766989708,
161          0.029504213482141495,
162          0.026794718578457832,
163          0.030350370332598686,
164          0.027094533666968346,
165          0.020488621667027473
166        ],
167        "n": 8
168      }
169    },
170    {
171      "cfg": {
172        "lr": 0.003
173      },
174      "full": {
175        "mean": 0.024627438746392727,
176        "std": 0.004374673952637387,
177        "per_seed": [
178          0.01918051764369011,
179          0.025951329618692398,
180          0.03225667402148247,
181          0.030034830793738365,
182          0.02338957041501999,
183          0.020098846405744553,
184          0.024956857785582542,
185          0.02115088328719139
186        ],
187        "n": 8
188      }
189    },
190    {
191      "cfg": {
192        "lr": 0.01
193      },
194      "full": {
195        "mean": 0.021506645949557424,
196        "std": 0.004870355712475741,
197        "per_seed": [
198          0.014883503317832947,
199          0.020688023418188095,
200          0.030457986518740654,
201          0.023641860112547874,
202          0.021189182996749878,
203          0.022041697055101395,
204          0.024708442389965057,
205          0.014442471787333488
206        ],
207        "n": 8
208      }
209    }
210  ],
211  "best_idea_cfg": {
212    "lr": 0.01
213  },
214  "mechanism_signature": {
215    "raw_mean_normal_component": 0.6702553331851959,
216    "projected_max_normal_component": 4.76837158203125e-07,
217    "rotation_relative_prediction_change": 0.25490777008235455,
218    "predicted_projected_leakage": "<1e-5",
219    "confirmed": true
220  }
221}