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": 0.005,
 10      "weight_decay": 0.0001
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "weight_decay": 0.0
 17        },
 18        "mean": 0.0014746972592547536
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.001,
 23          "weight_decay": 0.0001
 24        },
 25        "mean": 0.0014579330163542181
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.003,
 30          "weight_decay": 0.0
 31        },
 32        "mean": 0.0009116442524828017
 33      },
 34      {
 35        "cfg": {
 36          "lr": 0.003,
 37          "weight_decay": 0.0001
 38        },
 39        "mean": 0.000888297741767019
 40      },
 41      {
 42        "cfg": {
 43          "lr": 0.005,
 44          "weight_decay": 0.0
 45        },
 46        "mean": 0.000684085811371915
 47      },
 48      {
 49        "cfg": {
 50          "lr": 0.005,
 51          "weight_decay": 0.0001
 52        },
 53        "mean": 0.0006692118913633749
 54      }
 55    ],
 56    "full": {
 57      "mean": 0.0006622805813094601,
 58      "std": 0.00017074866873521193,
 59      "per_seed": [
 60        0.0006129607791081071,
 61        0.0005905718426220119,
 62        0.0005136265535838902,
 63        0.0009596883901394904,
 64        0.0005064630531705916,
 65        0.0005161517765372992,
 66        0.0006731264875270426,
 67        0.0009256557677872479
 68      ],
 69      "n": 8
 70    }
 71  },
 72  "idea": {
 73    "mean": 0.0007774294572300278,
 74    "std": 0.00018976167730898186,
 75    "per_seed": [
 76      0.0007154284394346178,
 77      0.0007136478670872748,
 78      0.0007075121975503862,
 79      0.0011022123508155346,
 80      0.000592997414059937,
 81      0.0005738030886277556,
 82      0.0007325921906158328,
 83      0.0010812421096488833
 84    ],
 85    "n": 8
 86  },
 87  "comparison": {
 88    "delta_mean": 0.00011514887592056766,
 89    "idea_wins": 0,
 90    "n_pairs": 8,
 91    "per_seed_diffs": [
 92      0.00010246766032651067,
 93      0.0001230760244652629,
 94      0.000193885643966496,
 95      0.00014252396067604423,
 96      8.653436088934541e-05,
 97      5.7651312090456486e-05,
 98      5.946570308879018e-05,
 99      0.00015558634186163545
100    ],
101    "p_value": 0.0081,
102    "mde": 4.0017114274026e-05,
103    "mde_rel_pct": 6.042320340255821,
104    "verdict": "idea worse (significant)",
105    "system_worked": false
106  },
107  "mechanism_signature": {
108    "prediction": "dissipativity penalty lowers trained local perturbation gain",
109    "baseline_mean_local_gain": 0.42203639075160027,
110    "idea_mean_local_gain": 0.4056360647082329,
111    "relative_reduction": 0.0388599808044046,
112    "per_seed": [
113      {
114        "seed": 0,
115        "baseline_gain": 0.40170010924339294,
116        "idea_gain": 0.39243876934051514
117      },
118      {
119        "seed": 1,
120        "baseline_gain": 0.4355578124523163,
121        "idea_gain": 0.40910080075263977
122      },
123      {
124        "seed": 2,
125        "baseline_gain": 0.4409540295600891,
126        "idea_gain": 0.41530841588974
127      },
128      {
129        "seed": 3,
130        "baseline_gain": 0.4010472893714905,
131        "idea_gain": 0.3920332193374634
132      },
133      {
134        "seed": 4,
135        "baseline_gain": 0.42631199955940247,
136        "idea_gain": 0.4098639488220215
137      },
138      {
139        "seed": 5,
140        "baseline_gain": 0.4153302311897278,
141        "idea_gain": 0.4018923342227936
142      },
143      {
144        "seed": 6,
145        "baseline_gain": 0.4280075430870056,
146        "idea_gain": 0.4121930003166199
147      },
148      {
149        "seed": 7,
150        "baseline_gain": 0.42738211154937744,
151        "idea_gain": 0.4122580289840698
152      }
153    ],
154    "confirmed": true
155  },
156  "idea_sweep": [
157    {
158      "lr": 0.005,
159      "result": {
160        "mean": 0.0007774294572300278,
161        "std": 0.00018976167730898186,
162        "per_seed": [
163          0.0007154284394346178,
164          0.0007136478670872748,
165          0.0007075121975503862,
166          0.0011022123508155346,
167          0.000592997414059937,
168          0.0005738030886277556,
169          0.0007325921906158328,
170          0.0010812421096488833
171        ],
172        "n": 8
173      }
174    },
175    {
176      "lr": 0.001,
177      "result": {
178        "mean": 0.0016683561116224155,
179        "std": 0.0005233230853334032,
180        "per_seed": [
181          0.0012873527593910694,
182          0.001405279035679996,
183          0.0013610285241156816,
184          0.001857919036410749,
185          0.0012282896786928177,
186          0.001367267337627709,
187          0.0019549040589481592,
188          0.002884808462113142
189        ],
190        "n": 8
191      }
192    },
193    {
194      "lr": 0.003,
195      "result": {
196        "mean": 0.0009767467927304097,
197        "std": 0.00019101000978935268,
198        "per_seed": [
199          0.0007808486698195338,
200          0.0008321516797877848,
201          0.0010572962928563356,
202          0.0012427946785464883,
203          0.0007201741682365537,
204          0.0008802887750789523,
205          0.0010501790093258023,
206          0.0012502410681918263
207        ],
208        "n": 8
209      }
210    }
211  ],
212  "protocol_notes": {
213    "matched_structure": "dynamics pendulum rollout; stability/control is the target domain",
214    "shared_architecture": "bench rnn_small GRU; only training penalty differs",
215    "shared_lr_union": [
216      0.001,
217      0.003,
218      0.005
219    ],
220    "epochs": 20,
221    "batch": 128,
222    "penalty": "relu(||d output/d input||_2^2 - target_gain^2)^2",
223    "target_gain": 0.35,
224    "lambda": 0.15
225  }
226}