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      "lam": 0.01,
 11      "epochs": 10
 12    },
 13    "sweep": [
 14      {
 15        "cfg": {
 16          "lr": 0.001,
 17          "lam": 0.001,
 18          "epochs": 10
 19        },
 20        "mean": 0.0017361862410325557
 21      },
 22      {
 23        "cfg": {
 24          "lr": 0.001,
 25          "lam": 0.01,
 26          "epochs": 10
 27        },
 28        "mean": 0.0017254647973459214
 29      },
 30      {
 31        "cfg": {
 32          "lr": 0.003,
 33          "lam": 0.001,
 34          "epochs": 10
 35        },
 36        "mean": 0.0011354178132023662
 37      },
 38      {
 39        "cfg": {
 40          "lr": 0.003,
 41          "lam": 0.01,
 42          "epochs": 10
 43        },
 44        "mean": 0.0011304349318379536
 45      },
 46      {
 47        "cfg": {
 48          "lr": 0.01,
 49          "lam": 0.001,
 50          "epochs": 10
 51        },
 52        "mean": 0.0006082321124267764
 53      },
 54      {
 55        "cfg": {
 56          "lr": 0.01,
 57          "lam": 0.01,
 58          "epochs": 10
 59        },
 60        "mean": 0.0006044723413651809
 61      }
 62    ],
 63    "full": {
 64      "mean": 0.0005673739506164566,
 65      "std": 0.00016282064525448582,
 66      "per_seed": [
 67        0.00044074832112528384,
 68        0.0007072906009852886,
 69        0.0003835938696283847,
 70        0.0008862565737217665,
 71        0.00040746666491031647,
 72        0.00047172914491966367,
 73        0.0006375783705152571,
 74        0.0006043280591256917
 75      ],
 76      "n": 8
 77    },
 78    "method": "pointwise constant selected-coordinate log-Jacobian"
 79  },
 80  "idea": {
 81    "mean": 0.0003382966906428919,
 82    "std": 0.0002807304207947017,
 83    "per_seed": [
 84      0.00012169728142907843,
 85      0.0010539937065914273,
 86      0.00020155370293650776,
 87      0.00022586695558857173,
 88      0.00014386164548341185,
 89      0.0003415524843148887,
 90      0.000293777440674603,
 91      0.00032407030812464654
 92    ],
 93    "n": 8,
 94    "method": "cohomological potential residual"
 95  },
 96  "comparison": {
 97    "delta_mean": -0.00022907725997356465,
 98    "idea_wins": 7,
 99    "n_pairs": 8,
100    "per_seed_diffs": [
101      -0.0003190510396962054,
102      0.0003467031056061387,
103      -0.00018204016669187695,
104      -0.0006603896181331947,
105      -0.0002636050194269046,
106      -0.00013017666060477495,
107      -0.00034380092984065413,
108      -0.0002802577510010451
109    ],
110    "p_value": 0.07245,
111    "mde": 0.0002353602814500899,
112    "mde_rel_pct": 41.482391145093814,
113    "verdict": "no significant win",
114    "system_worked": false
115  },
116  "mechanism_signature": {
117    "mechanism_signature": {
118      "prediction": "finite-horizon residual per step remains bounded rather than accumulating linearly",
119      "observed": [
120        {
121          "k": 1,
122          "observed_std_R_over_k": 0.05008017644286156,
123          "observed_std_ell_minus_c": 0.051468078047037125
124        },
125        {
126          "k": 4,
127          "observed_std_R_over_k": 0.05144515261054039,
128          "observed_std_ell_minus_c": 0.052239760756492615
129        },
130        {
131          "k": 8,
132          "observed_std_R_over_k": 0.05219043791294098,
133          "observed_std_ell_minus_c": 0.05265187472105026
134        }
135      ],
136      "confirmed": true
137    },
138    "idea_config_sweep": [
139      {
140        "cfg": {
141          "lr": 0.01,
142          "lam": 0.01,
143          "epochs": 10
144        },
145        "result": {
146          "mean": 0.0003382966906428919,
147          "std": 0.0002807304207947017,
148          "per_seed": [
149            0.00012169728142907843,
150            0.0010539937065914273,
151            0.00020155370293650776,
152            0.00022586695558857173,
153            0.00014386164548341185,
154            0.0003415524843148887,
155            0.000293777440674603,
156            0.00032407030812464654
157          ],
158          "n": 8,
159          "method": "cohomological potential residual"
160        }
161      },
162      {
163        "cfg": {
164          "lr": 0.001,
165          "lam": 0.01,
166          "epochs": 10
167        },
168        "result": {
169          "mean": 0.003544457780662924,
170          "std": 0.0019988002357774577,
171          "per_seed": [
172            0.002059740712866187,
173            0.006557196378707886,
174            0.0014491339679807425,
175            0.0023036343045532703,
176            0.0019505396485328674,
177            0.00565938837826252,
178            0.0023294156417250633,
179            0.006046613212674856
180          ],
181          "n": 8
182        }
183      },
184      {
185        "cfg": {
186          "lr": 0.01,
187          "lam": 0.01,
188          "epochs": 10
189        },
190        "result": {
191          "mean": 0.0003382966906428919,
192          "std": 0.0002807304207947017,
193          "per_seed": [
194            0.00012169728142907843,
195            0.0010539937065914273,
196            0.00020155370293650776,
197            0.00022586695558857173,
198            0.00014386164548341185,
199            0.0003415524843148887,
200            0.000293777440674603,
201            0.00032407030812464654
202          ],
203          "n": 8
204        }
205      }
206    ],
207    "device": "cuda",
208    "primary_metric": "test MSE"
209  },
210  "protocol_notes": "8 paired seeds; baseline sweep uses 4 seeds then full reevaluation; shared lr/lambda union"
211}