Cross-Degree Certificate Against Recurrent Oscillation / bench_report.json

Failed on benchmark

Raw ⬇ ZIP
  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": 0.006,
 10      "weight_decay": 0.0
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "weight_decay": 0.0
 17        },
 18        "mean": 0.012346302974037826
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "weight_decay": 0.0
 24        },
 25        "mean": 0.004180165706202388
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.006,
 30          "weight_decay": 0.0
 31        },
 32        "mean": 0.0025389266083948314
 33      }
 34    ],
 35    "full": {
 36      "mean": 0.0024029802589211613,
 37      "std": 0.0007328567905589416,
 38      "per_seed": [
 39        0.001808110624551773,
 40        0.0038746704813092947,
 41        0.002130982931703329,
 42        0.002341942396014929,
 43        0.001473060343414545,
 44        0.001914751948788762,
 45        0.003167124465107918,
 46        0.0025131988804787397
 47      ],
 48      "n": 8
 49    }
 50  },
 51  "idea": {
 52    "mean": 0.005464809873956256,
 53    "std": 0.002061025522134347,
 54    "per_seed": [
 55      0.007386896759271622,
 56      0.007307976484298706,
 57      0.006752352695912123,
 58      0.0009749423479661345,
 59      0.0034697847440838814,
 60      0.006274467799812555,
 61      0.006125933025032282,
 62      0.005426125135272741
 63    ],
 64    "n": 8
 65  },
 66  "comparison": {
 67    "delta_mean": 0.0030618296150350943,
 68    "idea_wins": 1,
 69    "n_pairs": 8,
 70    "per_seed_diffs": [
 71      0.005578786134719849,
 72      0.0034333060029894114,
 73      0.004621369764208794,
 74      -0.0013670000480487943,
 75      0.0019967244006693363,
 76      0.004359715851023793,
 77      0.002958808559924364,
 78      0.0029129262547940016
 79    ],
 80    "p_value": 0.0152,
 81    "mde": 0.0017701285805840077,
 82    "mde_rel_pct": 73.66388358840376,
 83    "verdict": "idea worse (significant)",
 84    "system_worked": false
 85  },
 86  "mechanism_signature": {
 87    "prediction": "anti-oscillation penalty should reduce trained recurrent two-step similarity relative to adjacent change",
 88    "baseline_mean_two_to_one_ratio": 2.1946372059836357,
 89    "idea_mean_two_to_one_ratio": 2.845283087236918,
 90    "predicted_direction": "idea lower than baseline",
 91    "confirmed": false,
 92    "per_seed": [
 93      {
 94        "seed": 0,
 95        "baseline": {
 96          "observed_two_step_mse": 0.004077367950230837,
 97          "observed_adjacent_mse": 0.001899817492812872,
 98          "observed_two_to_one_ratio": 2.1461779901889475
 99        },
100        "idea": {
101          "observed_two_step_mse": 4.882162102148868e-05,
102          "observed_adjacent_mse": 1.8978958905790932e-05,
103          "observed_two_to_one_ratio": 2.5710530663479334
104        }
105      },
106      {
107        "seed": 1,
108        "baseline": {
109          "observed_two_step_mse": 0.0035950804594904184,
110          "observed_adjacent_mse": 0.0015977286966517568,
111          "observed_two_to_one_ratio": 2.250105393969814
112        },
113        "idea": {
114          "observed_two_step_mse": 3.846972686005756e-05,
115          "observed_adjacent_mse": 1.3402634067460895e-05,
116          "observed_two_to_one_ratio": 2.868170910096271
117        }
118      },
119      {
120        "seed": 2,
121        "baseline": {
122          "observed_two_step_mse": 0.0043192701414227486,
123          "observed_adjacent_mse": 0.0019986697006970644,
124          "observed_two_to_one_ratio": 2.161061694836021
125        },
126        "idea": {
127          "observed_two_step_mse": 3.958393062930554e-05,
128          "observed_adjacent_mse": 1.3548745300795417e-05,
129          "observed_two_to_one_ratio": 2.9194390595259114
130        }
131      },
132      {
133        "seed": 3,
134        "baseline": {
135          "observed_two_step_mse": 0.003866632701829076,
136          "observed_adjacent_mse": 0.0016693203942850232,
137          "observed_two_to_one_ratio": 2.316277661430325
138        },
139        "idea": {
140          "observed_two_step_mse": 3.4838722058339044e-05,
141          "observed_adjacent_mse": 1.1200490916962735e-05,
142          "observed_two_to_one_ratio": 3.1076892454035296
143        }
144      },
145      {
146        "seed": 4,
147        "baseline": {
148          "observed_two_step_mse": 0.0038868181873112917,
149          "observed_adjacent_mse": 0.0016667143208906054,
150          "observed_two_to_one_ratio": 2.332010242242335
151        },
152        "idea": {
153          "observed_two_step_mse": 3.9646634832024574e-05,
154          "observed_adjacent_mse": 1.4838369679637253e-05,
155          "observed_two_to_one_ratio": 2.6701001987036426
156        }
157      },
158      {
159        "seed": 5,
160        "baseline": {
161          "observed_two_step_mse": 0.0034053579438477755,
162          "observed_adjacent_mse": 0.001578447176143527,
163          "observed_two_to_one_ratio": 2.157396472527507
164        },
165        "idea": {
166          "observed_two_step_mse": 4.4292191887507215e-05,
167          "observed_adjacent_mse": 1.595341927895788e-05,
168          "observed_two_to_one_ratio": 2.774605559968646
169        }
170      },
171      {
172        "seed": 6,
173        "baseline": {
174          "observed_two_step_mse": 0.0034227895084768534,
175          "observed_adjacent_mse": 0.0015517541905865073,
176          "observed_two_to_one_ratio": 2.205740749297205
177        },
178        "idea": {
179          "observed_two_step_mse": 3.656304397736676e-05,
180          "observed_adjacent_mse": 1.259200507774949e-05,
181          "observed_two_to_one_ratio": 2.9013671833797035
182        }
183      },
184      {
185        "seed": 7,
186        "baseline": {
187          "observed_two_step_mse": 0.003476944053545594,
188          "observed_adjacent_mse": 0.0017486677970737219,
189          "observed_two_to_one_ratio": 1.9883274433769293
190        },
191        "idea": {
192          "observed_two_step_mse": 3.67501015716698e-05,
193          "observed_adjacent_mse": 1.2448339475668035e-05,
194          "observed_two_to_one_ratio": 2.949839474469707
195        }
196      }
197    ]
198  },
199  "idea_trials": [
200    {
201      "cfg": {
202        "lr": 0.001,
203        "weight_decay": 0.0,
204        "lambda_cycle": 0.01
205      },
206      "result": {
207        "mean": 0.008634913916466758,
208        "std": 0.007319620916715749,
209        "per_seed": [
210          0.0085679329931736,
211          0.024327190592885017,
212          0.0032044434919953346,
213          0.00580877298489213,
214          0.003978467080742121,
215          0.004719811026006937,
216          0.016555899754166603,
217          0.0019167934078723192
218        ],
219        "n": 8
220      }
221    },
222    {
223      "cfg": {
224        "lr": 0.003,
225        "weight_decay": 0.0,
226        "lambda_cycle": 0.03
227      },
228      "result": {
229        "mean": 0.01017151348059997,
230        "std": 0.0028825955800251195,
231        "per_seed": [
232          0.009243861772119999,
233          0.009108055382966995,
234          0.013492719270288944,
235          0.015129592269659042,
236          0.010461220517754555,
237          0.009117119945585728,
238          0.004917753394693136,
239          0.009901785291731358
240        ],
241        "n": 8
242      }
243    },
244    {
245      "cfg": {
246        "lr": 0.006,
247        "weight_decay": 0.0,
248        "lambda_cycle": 0.1
249      },
250      "result": {
251        "mean": 0.005464809873956256,
252        "std": 0.002061025522134347,
253        "per_seed": [
254          0.007386896759271622,
255          0.007307976484298706,
256          0.006752352695912123,
257          0.0009749423479661345,
258          0.0034697847440838814,
259          0.006274467799812555,
260          0.006125933025032282,
261          0.005426125135272741
262        ],
263        "n": 8
264      }
265    }
266  ],
267  "structural_match": "Dynamics track: recurrent GRU predicts actuated pendulum rollout; anti-oscillation targets recurrent hidden-state period-two behaviour.",
268  "protocol": {
269    "paired_seeds": [
270      0,
271      1,
272      2,
273      3,
274      4,
275      5,
276      6,
277      7
278    ],
279    "epochs": 12,
280    "batch": 128,
281    "baseline_grid": [
282      {
283        "lr": 0.001,
284        "weight_decay": 0.0
285      },
286      {
287        "lr": 0.003,
288        "weight_decay": 0.0
289      },
290      {
291        "lr": 0.006,
292        "weight_decay": 0.0
293      }
294    ],
295    "idea_grid": [
296      {
297        "lr": 0.001,
298        "weight_decay": 0.0,
299        "lambda_cycle": 0.01
300      },
301      {
302        "lr": 0.003,
303        "weight_decay": 0.0,
304        "lambda_cycle": 0.03
305      },
306      {
307        "lr": 0.006,
308        "weight_decay": 0.0,
309        "lambda_cycle": 0.1
310      }
311    ]
312  }
313}