1{
  2  "bench_version": 1,
  3  "track": "dynamics",
  4  "model": "rnn_small",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.01,
 10      "gate_mid": 0.08,
 11      "gate_slope": 3.0
 12    },
 13    "sweep": [
 14      {
 15        "cfg": {
 16          "lr": 0.001,
 17          "gate_mid": 0.08,
 18          "gate_slope": 3.0
 19        },
 20        "mean": 0.001094405582989566
 21      },
 22      {
 23        "cfg": {
 24          "lr": 0.003,
 25          "gate_mid": 0.08,
 26          "gate_slope": 3.0
 27        },
 28        "mean": 0.0005730643169954419
 29      },
 30      {
 31        "cfg": {
 32          "lr": 0.01,
 33          "gate_mid": 0.08,
 34          "gate_slope": 3.0
 35        },
 36        "mean": 0.0001522819948149845
 37      }
 38    ],
 39    "full": {
 40      "mean": 0.00012016432719974546,
 41      "std": 5.673950443795311e-05,
 42      "per_seed": [
 43        0.00023659199359826744,
 44        0.000133594949147664,
 45        7.917142647784203e-05,
 46        0.00015976961003616452,
 47        6.463151657953858e-05,
 48        9.197137842420489e-05,
 49        5.368743586586788e-05,
 50        0.0001418963074684143
 51      ],
 52      "n": 8
 53    }
 54  },
 55  "idea": {
 56    "mean": 4.0656192368260236e-06,
 57    "std": 1.5779659647239028e-07,
 58    "per_seed": [
 59      4.112095211894484e-06,
 60      3.850047960440861e-06,
 61      3.88751095670159e-06,
 62      4.318793344282312e-06,
 63      3.984665454481728e-06,
 64      3.9648107303946745e-06,
 65      4.202797754260246e-06,
 66      4.2042324821522925e-06
 67    ],
 68    "n": 8
 69  },
 70  "comparison": {
 71    "delta_mean": -0.00011609870796291943,
 72    "idea_wins": 8,
 73    "n_pairs": 8,
 74    "per_seed_diffs": [
 75      -0.00023247989838637295,
 76      -0.00012974490118722315,
 77      -7.528391552114044e-05,
 78      -0.0001554508166918822,
 79      -6.0646851125056855e-05,
 80      -8.800656769381021e-05,
 81      -4.948463811160764e-05,
 82      -0.00013769207498626201
 83    ],
 84    "p_value": 0.0081,
 85    "mde": 5.06773245421868e-05,
 86    "mde_rel_pct": 42.17335187833861,
 87    "verdict": "idea better (significant)",
 88    "system_worked": true
 89  },
 90  "mechanism_signature": {
 91    "trained_model": true,
 92    "defect_low_closure": 0.0020928073208779097,
 93    "defect_high_closure": 0.01778038963675499,
 94    "defect_closure_corr": 0.9902236612013674,
 95    "background_rmse": 0.008130352012813091,
 96    "prediction": "closure should rise with defect and fall as background becomes accurate",
 97    "confirmed": true
 98  },
 99  "idea_sweep": [
100    {
101      "cfg": {
102        "lr": 0.001,
103        "gate_mid": 0.08,
104        "gate_slope": 3.0
105      },
106      "result": {
107        "mean": 4.137774908485881e-06,
108        "std": 1.5718023689435836e-07,
109        "per_seed": [
110          4.205145160085522e-06,
111          3.896001544489991e-06,
112          4.057551450387109e-06,
113          4.373848241812084e-06,
114          4.077617177244974e-06,
115          3.985380772064673e-06,
116          4.350635663286084e-06,
117          4.156019258516608e-06
118        ],
119        "n": 8
120      }
121    },
122    {
123      "cfg": {
124        "lr": 0.003,
125        "gate_mid": 0.08,
126        "gate_slope": 3.0
127      },
128      "result": {
129        "mean": 4.0656192368260236e-06,
130        "std": 1.5779659647239028e-07,
131        "per_seed": [
132          4.112095211894484e-06,
133          3.850047960440861e-06,
134          3.88751095670159e-06,
135          4.318793344282312e-06,
136          3.984665454481728e-06,
137          3.9648107303946745e-06,
138          4.202797754260246e-06,
139          4.2042324821522925e-06
140        ],
141        "n": 8
142      }
143    },
144    {
145      "cfg": {
146        "lr": 0.01,
147        "gate_mid": 0.08,
148        "gate_slope": 3.0
149      },
150      "result": {
151        "mean": 4.102830018837267e-06,
152        "std": 1.5946786154305604e-07,
153        "per_seed": [
154          4.118781362194568e-06,
155          3.881638804159593e-06,
156          3.890218522428768e-06,
157          4.2627821130736265e-06,
158          4.292129688110435e-06,
159          3.9507722249254584e-06,
160          4.203915068501374e-06,
161          4.222402367304312e-06
162        ],
163        "n": 8
164      }
165    }
166  ],
167  "protocol": {
168    "paired_seeds": [
169      0,
170      1,
171      2,
172      3,
173      4,
174      5,
175      6,
176      7
177    ],
178    "sweep_seeds": [
179      0,
180      1,
181      2,
182      3
183    ],
184    "epochs": 15,
185    "n_train": 1000,
186    "n_test": 400,
187    "same_architecture": true,
188    "baseline_grid": [
189      {
190        "lr": 0.001,
191        "gate_mid": 0.08,
192        "gate_slope": 3.0
193      },
194      {
195        "lr": 0.003,
196        "gate_mid": 0.08,
197        "gate_slope": 3.0
198      },
199      {
200        "lr": 0.01,
201        "gate_mid": 0.08,
202        "gate_slope": 3.0
203      }
204    ],
205    "idea_grid": [
206      {
207        "lr": 0.001,
208        "gate_mid": 0.08,
209        "gate_slope": 3.0
210      },
211      {
212        "lr": 0.003,
213        "gate_mid": 0.08,
214        "gate_slope": 3.0
215      },
216      {
217        "lr": 0.01,
218        "gate_mid": 0.08,
219        "gate_slope": 3.0
220      }
221    ],
222    "intervention": "analytic pendulum background, residual coordinates, defect sigmoid gate"
223  }
224}