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    },
 11    "sweep": [
 12      {
 13        "cfg": {
 14          "lr": 0.001
 15        },
 16        "mean": 0.0013811723329126835
 17      },
 18      {
 19        "cfg": {
 20          "lr": 0.003
 21        },
 22        "mean": 0.0009251563169527799
 23      },
 24      {
 25        "cfg": {
 26          "lr": 0.01
 27        },
 28        "mean": 0.00036846242801402695
 29      }
 30    ],
 31    "full": {
 32      "mean": 0.00038149026113387663,
 33      "std": 9.695924201469225e-05,
 34      "per_seed": [
 35        0.00029065465787425637,
 36        0.0005194233381189406,
 37        0.0002394697949057445,
 38        0.00042430192115716636,
 39        0.00026041336241178215,
 40        0.00043081643525511026,
 41        0.0004192772030364722,
 42        0.0004675653763115406
 43      ],
 44      "n": 8
 45    }
 46  },
 47  "idea": {
 48    "mean": 0.000929887879465241,
 49    "std": 0.00022563265805066197,
 50    "per_seed": [
 51      0.0007356178830377758,
 52      0.0008692095289006829,
 53      0.0007401101174764335,
 54      0.0013556877383962274,
 55      0.0007141538080759346,
 56      0.0007827341323718429,
 57      0.0010597598738968372,
 58      0.0011818299535661936
 59    ],
 60    "n": 8
 61  },
 62  "comparison": {
 63    "delta_mean": 0.0005483976183313644,
 64    "idea_wins": 0,
 65    "n_pairs": 8,
 66    "per_seed_diffs": [
 67      0.0004449632251635194,
 68      0.00034978619078174233,
 69      0.000500640322570689,
 70      0.000931385817239061,
 71      0.0004537404456641525,
 72      0.0003519176971167326,
 73      0.000640482670860365,
 74      0.000714264577254653
 75    ],
 76    "p_value": 0.0081,
 77    "mde": 0.00016802373854812774,
 78    "mde_rel_pct": 44.04404402060556,
 79    "verdict": "idea worse (significant)",
 80    "system_worked": false
 81  },
 82  "mechanism_signature": {
 83    "mechanism_signature": {
 84      "prediction": "beta*r approximately constant across delayed bursts",
 85      "observed_n": 0,
 86      "observed_beta_r": [],
 87      "mean_beta_r": null,
 88      "relative_spread": null,
 89      "trained_model_escape_events": 0,
 90      "confirmed": false
 91    },
 92    "idea_sweep": [
 93      {
 94        "cfg": {
 95          "lr": 0.01,
 96          "delay": 2
 97        },
 98        "mean": 0.002265104980324395
 99      },
100      {
101        "cfg": {
102          "lr": 0.003,
103          "delay": 1
104        },
105        "mean": 0.000929887879465241
106      },
107      {
108        "cfg": {
109          "lr": 0.01,
110          "delay": 4
111        },
112        "mean": 0.008774380781687796
113      }
114    ]
115  },
116  "selected_idea_cfg": {
117    "lr": 0.003,
118    "delay": 1
119  },
120  "seed_details": [
121    {
122      "loss": [
123        0.5282221913337708,
124        0.08002735763788223,
125        0.040321074873209,
126        0.027806747257709503,
127        0.00558410506695509,
128        0.00680130623281002,
129        0.0032351754978299143,
130        0.001620920547284186,
131        0.001523228024598211,
132        0.0011141636362299323,
133        0.000860015437938273,
134        0.0008175975293852389
135      ],
136      "escape_times": [],
137      "n": 0,
138      "r": null,
139      "beta": null,
140      "beta_r": null,
141      "ks_p": null,
142      "gumbel_aic": null,
143      "exp_aic": null,
144      "halted_bursts": 0,
145      "q_start": null,
146      "q_end": null
147    },
148    {
149      "loss": [
150        0.5754758775234222,
151        0.12694327160716057,
152        0.056237182226032016,
153        0.02086150236427784,
154        0.011539126932621002,
155        0.008746378906071186,
156        0.0024228729121387005,
157        0.002637408571317792,
158        0.0015756288496777415,
159        0.000965956358704716,
160        0.0009452591021545232,
161        0.0008290903922170401
162      ],
163      "escape_times": [],
164      "n": 0,
165      "r": null,
166      "beta": null,
167      "beta_r": null,
168      "ks_p": null,
169      "gumbel_aic": null,
170      "exp_aic": null,
171      "halted_bursts": 0,
172      "q_start": null,
173      "q_end": null
174    },
175    {
176      "loss": [
177        0.4696029615402222,
178        0.09072692170739174,
179        0.045659169703722,
180        0.017878399193286897,
181        0.012400263845920563,
182        0.003765859454870224,
183        0.003672598684206605,
184        0.0014206368429586292,
185        0.0009925120219122618,
186        0.0008811178850010037,
187        0.0006947765126824379,
188        0.0005901444423943758
189      ],
190      "escape_times": [],
191      "n": 0,
192      "r": null,
193      "beta": null,
194      "beta_r": null,
195      "ks_p": null,
196      "gumbel_aic": null,
197      "exp_aic": null,
198      "halted_bursts": 0,
199      "q_start": null,
200      "q_end": null
201    },
202    {
203      "loss": [
204        0.5658795094490051,
205        0.13714899322018026,
206        0.045997631624341014,
207        0.01840634111315012,
208        0.007587166950106621,
209        0.006993172634392977,
210        0.0028821277152746914,
211        0.00157423485070467,
212        0.0014174694032408298,
213        0.001179328037542291,
214        0.0008535732934251427,
215        0.0007487329898867756
216      ],
217      "escape_times": [],
218      "n": 0,
219      "r": null,
220      "beta": null,
221      "beta_r": null,
222      "ks_p": null,
223      "gumbel_aic": null,
224      "exp_aic": null,
225      "halted_bursts": 0,
226      "q_start": null,
227      "q_end": null
228    },
229    {
230      "loss": [
231        0.44831172585487367,
232        0.05728273868560791,
233        0.03125433370471001,
234        0.02487820740789175,
235        0.004413171038031578,
236        0.005216334542492405,
237        0.003302942952141166,
238        0.0015337833855301143,
239        0.0015170708391815423,
240        0.0010195694025605917,
241        0.000875311573036015,
242        0.0007151975075248628
243      ],
244      "escape_times": [],
245      "n": 0,
246      "r": null,
247      "beta": null,
248      "beta_r": null,
249      "ks_p": null,
250      "gumbel_aic": null,
251      "exp_aic": null,
252      "halted_bursts": 0,
253      "q_start": null,
254      "q_end": null
255    },
256    {
257      "loss": [
258        0.5917379081249237,
259        0.1275447541475296,
260        0.05130602061748504,
261        0.028465592414140702,
262        0.00793339565396309,
263        0.008613825049251319,
264        0.0035424450412392615,
265        0.0028161489311605693,
266        0.0022506765369325874,
267        0.0014946001255884767,
268        0.0013238138472661375,
269        0.0011464648973196745
270      ],
271      "escape_times": [],
272      "n": 0,
273      "r": null,
274      "beta": null,
275      "beta_r": null,
276      "ks_p": null,
277      "gumbel_aic": null,
278      "exp_aic": null,
279      "halted_bursts": 0,
280      "q_start": null,
281      "q_end": null
282    },
283    {
284      "loss": [
285        0.5767341089248658,
286        0.1493582247197628,
287        0.05171388104557991,
288        0.013434727936983108,
289        0.014380061002448201,
290        0.004003392308950424,
291        0.0036623673141002655,
292        0.0027428057603538034,
293        0.0016886063292622566,
294        0.001497098971158266,
295        0.0011919403914362192,
296        0.0011099096620455385
297      ],
298      "escape_times": [],
299      "n": 0,
300      "r": null,
301      "beta": null,
302      "beta_r": null,
303      "ks_p": null,
304      "gumbel_aic": null,
305      "exp_aic": null,
306      "halted_bursts": 0,
307      "q_start": null,
308      "q_end": null
309    },
310    {
311      "loss": [
312        0.5823761320114136,
313        0.15505388271063567,
314        0.04796165257692337,
315        0.012322895303368568,
316        0.015082798413932324,
317        0.004147060438990593,
318        0.004272648510523141,
319        0.002224865108728409,
320        0.001488620461896062,
321        0.0013844934199005365,
322        0.000926615516655147,
323        0.000944888039957732
324      ],
325      "escape_times": [],
326      "n": 0,
327      "r": null,
328      "beta": null,
329      "beta_r": null,
330      "ks_p": null,
331      "gumbel_aic": null,
332      "exp_aic": null,
333      "halted_bursts": 0,
334      "q_start": null,
335      "q_end": null
336    }
337  ]
338}