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": 0.003
 10    },
 11    "sweep": [
 12      {
 13        "cfg": {
 14          "lr": 0.001
 15        },
 16        "mean": 0.010543678537942469
 17      },
 18      {
 19        "cfg": {
 20          "lr": 0.003
 21        },
 22        "mean": 0.008196698385290802
 23      },
 24      {
 25        "cfg": {
 26          "lr": 0.01
 27        },
 28        "mean": 0.009198607411235571
 29      }
 30    ],
 31    "full": {
 32      "mean": 0.007723776740022004,
 33      "std": 0.0016993505269699558,
 34      "per_seed": [
 35        0.0053276014514267445,
 36        0.00833857711404562,
 37        0.008281959220767021,
 38        0.01083865575492382,
 39        0.008155343122780323,
 40        0.005130748730152845,
 41        0.007964092306792736,
 42        0.007753236219286919
 43      ],
 44      "n": 8
 45    }
 46  },
 47  "idea": {
 48    "mean": 0.008810116676613688,
 49    "std": 0.0023181915418372715,
 50    "per_seed": [
 51      0.009634214453399181,
 52      0.007696767803281546,
 53      0.01108845230191946,
 54      0.013114262372255325,
 55      0.009547872468829155,
 56      0.006006436888128519,
 57      0.006788569502532482,
 58      0.006604357622563839
 59    ],
 60    "n": 8
 61  },
 62  "comparison": {
 63    "delta_mean": 0.0010863399365916848,
 64    "idea_wins": 3,
 65    "n_pairs": 8,
 66    "per_seed_diffs": [
 67      0.004306613001972437,
 68      -0.0006418093107640743,
 69      0.002806493081152439,
 70      0.002275606617331505,
 71      0.001392529346048832,
 72      0.0008756881579756737,
 73      -0.001175522804260254,
 74      -0.0011488785967230797
 75    ],
 76    "p_value": 0.1829,
 77    "mde": 0.0016717188889132512,
 78    "mde_rel_pct": 21.643801279896767,
 79    "verdict": "no significant win",
 80    "system_worked": false,
 81    "idea_sweep_results": [
 82      {
 83        "cfg": {
 84          "lr": 0.003,
 85          "steps": 8,
 86          "step_size": 0.2
 87        },
 88        "result": {
 89          "mean": 0.010117848869413137,
 90          "std": 0.002055095696012117,
 91          "per_seed": [
 92            0.006481844000518322,
 93            0.009532518684864044,
 94            0.011807672679424286,
 95            0.014011072926223278,
 96            0.009151139296591282,
 97            0.010334501974284649,
 98            0.010534162633121014,
 99            0.009089878760278225
100          ],
101          "n": 8
102        }
103      },
104      {
105        "cfg": {
106          "lr": 0.001,
107          "steps": 8,
108          "step_size": 0.2
109        },
110        "result": {
111          "mean": 0.011166991549544036,
112          "std": 0.0018518637822724081,
113          "per_seed": [
114            0.008419685997068882,
115            0.009691255167126656,
116            0.01094791665673256,
117            0.013981804251670837,
118            0.012324392795562744,
119            0.011532868258655071,
120            0.013251206837594509,
121            0.009186802431941032
122          ],
123          "n": 8
124        }
125      },
126      {
127        "cfg": {
128          "lr": 0.01,
129          "steps": 8,
130          "step_size": 0.2
131        },
132        "result": {
133          "mean": 0.008810116676613688,
134          "std": 0.0023181915418372715,
135          "per_seed": [
136            0.009634214453399181,
137            0.007696767803281546,
138            0.01108845230191946,
139            0.013114262372255325,
140            0.009547872468829155,
141            0.006006436888128519,
142            0.006788569502532482,
143            0.006604357622563839
144          ],
145          "n": 8
146        }
147      }
148    ]
149  },
150  "track_justification": "Dynamics is structurally matched: the built-in task is an actuated pendulum rollout whose target depends on coupled state-transition mechanisms.",
151  "baseline_sweep_union_lr": [
152    0.001,
153    0.003,
154    0.01
155  ],
156  "idea_sweep": [
157    {
158      "cfg": {
159        "lr": 0.003,
160        "steps": 8,
161        "step_size": 0.2
162      },
163      "result": {
164        "mean": 0.010117848869413137,
165        "std": 0.002055095696012117,
166        "per_seed": [
167          0.006481844000518322,
168          0.009532518684864044,
169          0.011807672679424286,
170          0.014011072926223278,
171          0.009151139296591282,
172          0.010334501974284649,
173          0.010534162633121014,
174          0.009089878760278225
175        ],
176        "n": 8
177      }
178    },
179    {
180      "cfg": {
181        "lr": 0.001,
182        "steps": 8,
183        "step_size": 0.2
184      },
185      "result": {
186        "mean": 0.011166991549544036,
187        "std": 0.0018518637822724081,
188        "per_seed": [
189          0.008419685997068882,
190          0.009691255167126656,
191          0.01094791665673256,
192          0.013981804251670837,
193          0.012324392795562744,
194          0.011532868258655071,
195          0.013251206837594509,
196          0.009186802431941032
197        ],
198        "n": 8
199      }
200    },
201    {
202      "cfg": {
203        "lr": 0.01,
204        "steps": 8,
205        "step_size": 0.2
206      },
207      "result": {
208        "mean": 0.008810116676613688,
209        "std": 0.0023181915418372715,
210        "per_seed": [
211          0.009634214453399181,
212          0.007696767803281546,
213          0.01108845230191946,
214          0.013114262372255325,
215          0.009547872468829155,
216          0.006006436888128519,
217          0.006788569502532482,
218          0.006604357622563839
219        ],
220        "n": 8
221      }
222    }
223  ],
224  "idea_selected_cfg": {
225    "lr": 0.01,
226    "steps": 8,
227    "step_size": 0.2
228  },
229  "parameter_counts": {
230    "baseline": 3601,
231    "idea": 11043
232  },
233  "mechanism_signature": {
234    "prediction": "same xi imposed by different equation replacements yields different downstream state",
235    "n_samples": 64,
236    "xi": 0.08397173136472702,
237    "trained_model_target_error_f0": 0.44350090622901917,
238    "trained_model_target_error_f1": 0.44543659687042236,
239    "trained_model_downstream_separation": 0.2748295068740845,
240    "baseline_vs_observational_prediction_abs_gap": 0.05360101908445358,
241    "confirmed": false
242  }
243}