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": 0.002,
 10      "weight_decay": 0.0001
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.0005,
 16          "weight_decay": 0.0
 17        },
 18        "mean": 0.11454633437097073
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.0005,
 23          "weight_decay": 0.0001
 24        },
 25        "mean": 0.11465956829488277
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.001,
 30          "weight_decay": 0.0
 31        },
 32        "mean": 0.005499854683876038
 33      },
 34      {
 35        "cfg": {
 36          "lr": 0.001,
 37          "weight_decay": 0.0001
 38        },
 39        "mean": 0.005453515565022826
 40      },
 41      {
 42        "cfg": {
 43          "lr": 0.002,
 44          "weight_decay": 0.0
 45        },
 46        "mean": 0.004555012041237205
 47      },
 48      {
 49        "cfg": {
 50          "lr": 0.002,
 51          "weight_decay": 0.0001
 52        },
 53        "mean": 0.004531972575932741
 54      }
 55    ],
 56    "full": {
 57      "mean": 0.004597172053763643,
 58      "std": 0.0008625188036902094,
 59      "per_seed": [
 60        0.004187586717307568,
 61        0.006327331997454166,
 62        0.0038855248130857944,
 63        0.003727446775883436,
 64        0.005093815270811319,
 65        0.005022076889872551,
 66        0.004930851981043816,
 67        0.0036027419846504927
 68      ],
 69      "n": 8
 70    },
 71    "all_union_configs": [
 72      {
 73        "lr": 0.0005,
 74        "weight_decay": 0.0
 75      },
 76      {
 77        "lr": 0.0005,
 78        "weight_decay": 0.0001
 79      },
 80      {
 81        "lr": 0.001,
 82        "weight_decay": 0.0
 83      },
 84      {
 85        "lr": 0.001,
 86        "weight_decay": 0.0001
 87      },
 88      {
 89        "lr": 0.002,
 90        "weight_decay": 0.0
 91      },
 92      {
 93        "lr": 0.002,
 94        "weight_decay": 0.0001
 95      }
 96    ]
 97  },
 98  "idea": {
 99    "mean": 0.7925681881606579,
100    "std": 0.47593743937124783,
101    "per_seed": [
102      0.9278137683868408,
103      1.54276442527771,
104      0.2908315360546112,
105      1.384081244468689,
106      0.39941179752349854,
107      0.5426146984100342,
108      1.049137830734253,
109      0.20389020442962646
110    ],
111    "n": 8
112  },
113  "comparison": {
114    "delta_mean": 0.7879710161068942,
115    "idea_wins": 0,
116    "n_pairs": 8,
117    "per_seed_diffs": [
118      0.9236261816695333,
119      1.5364370932802558,
120      0.2869460112415254,
121      1.3803537976928055,
122      0.3943179822526872,
123      0.5375926215201616,
124      1.0442069787532091,
125      0.20028746244497597
126    ],
127    "p_value": 0.0081,
128    "mde": 0.4250928421498136,
129    "mde_rel_pct": 9246.833426688821,
130    "verdict": "idea worse (significant)",
131    "system_worked": false
132  },
133  "mechanism_signature": {
134    "quantity": "sigma(theta) at ||theta||=sqrt(d), measured formula in trained-update state space",
135    "predicted_slope": 0.6931471805599453,
136    "observed_slope": 0.6931471805599447,
137    "max_abs_error": 0.0,
138    "confirmed": true,
139    "note": "The update mechanism was exercised by trained dynamics-track models; this signature tests coefficient scaling, not sampling ESS."
140  },
141  "idea_sweep": [
142    {
143      "cfg": {
144        "lr": 0.002,
145        "weight_decay": 0.0001,
146        "alpha": 0.05
147      },
148      "full": {
149        "mean": 2.3324266374111176,
150        "std": 0.4043937352445423,
151        "per_seed": [
152          2.4503955841064453,
153          2.867910385131836,
154          2.2843286991119385,
155          1.803705096244812,
156          2.6125521659851074,
157          2.8163137435913086,
158          2.085012912750244,
159          1.7391945123672485
160        ],
161        "n": 8
162      }
163    },
164    {
165      "cfg": {
166        "lr": 0.0005,
167        "weight_decay": 0.0001,
168        "alpha": 0.05
169      },
170      "full": {
171        "mean": 0.7925681881606579,
172        "std": 0.47593743937124783,
173        "per_seed": [
174          0.9278137683868408,
175          1.54276442527771,
176          0.2908315360546112,
177          1.384081244468689,
178          0.39941179752349854,
179          0.5426146984100342,
180          1.049137830734253,
181          0.20389020442962646
182        ],
183        "n": 8
184      }
185    },
186    {
187      "cfg": {
188        "lr": 0.002,
189        "weight_decay": 0.0001,
190        "alpha": 0.05
191      },
192      "full": {
193        "mean": 2.3324266374111176,
194        "std": 0.4043937352445423,
195        "per_seed": [
196          2.4503955841064453,
197          2.867910385131836,
198          2.2843286991119385,
199          1.803705096244812,
200          2.6125521659851074,
201          2.8163137435913086,
202          2.085012912750244,
203          1.7391945123672485
204        ],
205        "n": 8
206      }
207    }
208  ],
209  "transfer_note": "No built-in latent/energy sampler track exists; dynamics is the closest structural stability/control track. The intervention is therefore a parameter-space Langevin transfer, not latent-state sampling."
210}