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": 0.012,
 10      "momentum": 0.9
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.003,
 16          "momentum": 0.8
 17        },
 18        "mean": 0.0007268761837622151
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "momentum": 0.9
 24        },
 25        "mean": 0.00030221821725717746
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.006,
 30          "momentum": 0.8
 31        },
 32        "mean": 0.00032218691558227874
 33      },
 34      {
 35        "cfg": {
 36          "lr": 0.006,
 37          "momentum": 0.9
 38        },
 39        "mean": 0.00023875456099631265
 40      },
 41      {
 42        "cfg": {
 43          "lr": 0.012,
 44          "momentum": 0.8
 45        },
 46        "mean": 0.0002535910825827159
 47      },
 48      {
 49        "cfg": {
 50          "lr": 0.012,
 51          "momentum": 0.9
 52        },
 53        "mean": 0.00020586683422152419
 54      }
 55    ],
 56    "full": {
 57      "mean": 0.00022353084659698652,
 58      "std": 9.061581347536935e-05,
 59      "per_seed": [
 60        0.00010926207323791459,
 61        0.00024170543474610895,
 62        0.0001626280864002183,
 63        0.0003098717425018549,
 64        0.0001496187032898888,
 65        0.00020368325931485742,
 66        0.0004100580408703536,
 67        0.00020141943241469562
 68      ],
 69      "n": 8
 70    },
 71    "parity_grid": [
 72      {
 73        "lr": 0.003,
 74        "momentum": 0.8
 75      },
 76      {
 77        "lr": 0.003,
 78        "momentum": 0.9
 79      },
 80      {
 81        "lr": 0.006,
 82        "momentum": 0.8
 83      },
 84      {
 85        "lr": 0.006,
 86        "momentum": 0.9
 87      },
 88      {
 89        "lr": 0.012,
 90        "momentum": 0.8
 91      },
 92      {
 93        "lr": 0.012,
 94        "momentum": 0.9
 95      }
 96    ]
 97  },
 98  "idea": {
 99    "mean": 0.000214986320315802,
100    "std": 8.761713194482366e-05,
101    "per_seed": [
102      0.00010610528261167929,
103      0.00024355029745493084,
104      0.00016700058768037707,
105      0.0002708751999307424,
106      0.0001304265606449917,
107      0.0002128796768374741,
108      0.000402944308007136,
109      0.00018610864935908467
110    ],
111    "n": 8
112  },
113  "comparison": {
114    "delta_mean": -8.544526281184517e-06,
115    "idea_wins": 5,
116    "n_pairs": 8,
117    "per_seed_diffs": [
118      -3.1567906262353063e-06,
119      1.8448627088218927e-06,
120      4.372501280158758e-06,
121      -3.8996542571112514e-05,
122      -1.9192142644897103e-05,
123      9.196417522616684e-06,
124      -7.113732863217592e-06,
125      -1.5310783055610955e-05
126    ],
127    "p_value": 0.17475,
128    "mde": 1.3057637424010761e-05,
129    "mde_rel_pct": 5.841537140309294,
130    "verdict": "no significant win",
131    "system_worked": false
132  },
133  "mechanism_signature": {
134    "track_match": "stability/control -> dynamics",
135    "idea_sweep": [
136      {
137        "cfg": {
138          "lr": 0.003,
139          "momentum": 0.8,
140          "branch_ratio": 2.0
141        },
142        "mean": 0.0006308200972853228
143      },
144      {
145        "cfg": {
146          "lr": 0.006,
147          "momentum": 0.8,
148          "branch_ratio": 2.0
149        },
150        "mean": 0.00030224852162064053
151      },
152      {
153        "cfg": {
154          "lr": 0.012,
155          "momentum": 0.9,
156          "branch_ratio": 2.0
157        },
158        "mean": 0.0001968828419194324
159      }
160    ],
161    "best_idea_cfg": {
162      "lr": 0.012,
163      "momentum": 0.9,
164      "branch_ratio": 2.0
165    },
166    "predicted_branch": 0,
167    "predicted_drift": [
168      10.384744470593654,
169      20.779192115849302
170    ],
171    "observed_selected_fraction_branch0": 0.65,
172    "confirmed": true,
173    "definition": "NN-scale projected parameter radial drift; lower predicted drift branch should be selected",
174    "n_probe_updates": 120
175  }
176}