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      "cycle_weight": 0.0
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "cycle_weight": 0.0
 17        },
 18        "mean": 0.001864717691205442
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "cycle_weight": 0.0
 24        },
 25        "mean": 0.0010901302302954718
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.01,
 30          "cycle_weight": 0.0
 31        },
 32        "mean": 0.0006186682076076977
 33      }
 34    ],
 35    "full": {
 36      "mean": 0.00054517226817552,
 37      "std": 0.00020361842727135583,
 38      "per_seed": [
 39        0.0004930912400595844,
 40        0.0004765945195686072,
 41        0.0004673648509196937,
 42        0.0010376222198829055,
 43        0.000511663849465549,
 44        0.00027233464061282575,
 45        0.0005571767687797546,
 46        0.0005455300561152399
 47      ],
 48      "n": 8
 49    }
 50  },
 51  "idea": {
 52    "mean": 0.00045470128861779813,
 53    "std": 0.00014832940019044596,
 54    "per_seed": [
 55      0.0004892470315098763,
 56      0.0003560753830242902,
 57      0.0005445053684525192,
 58      0.0007670727791264653,
 59      0.00048597707063890994,
 60      0.00024241428764071316,
 61      0.0003992895071860403,
 62      0.0003530288813635707
 63    ],
 64    "n": 8
 65  },
 66  "comparison": {
 67    "delta_mean": -9.047097955772188e-05,
 68    "idea_wins": 7,
 69    "n_pairs": 8,
 70    "per_seed_diffs": [
 71      -3.844208549708128e-06,
 72      -0.00012051913654431701,
 73      7.714051753282547e-05,
 74      -0.00027054944075644016,
 75      -2.5686778826639056e-05,
 76      -2.9920352972112596e-05,
 77      -0.00015788726159371436,
 78      -0.00019250117475166917
 79    ],
 80    "p_value": 0.07265,
 81    "mde": 9.572692629336764e-05,
 82    "mde_rel_pct": 17.559023428269473,
 83    "verdict": "no significant win",
 84    "system_worked": false
 85  },
 86  "mechanism_signature": {
 87    "cfg": {
 88      "lr": 0.01,
 89      "cycle_weight": 0.03
 90    },
 91    "behavior": {
 92      "prediction": "small input perturbations preserve decoded joint state",
 93      "predicted": 1.0,
 94      "observed": 1.0,
 95      "tolerance": 0.1,
 96      "confirmed": true,
 97      "n_models": 8
 98    }
 99  },
100  "idea_sweep": [
101    {
102      "cfg": {
103        "lr": 0.01,
104        "cycle_weight": 0.03
105      },
106      "mean": 0.00045470128861779813,
107      "std": 0.00014832940019044596,
108      "per_seed": [
109        0.0004892470315098763,
110        0.0003560753830242902,
111        0.0005445053684525192,
112        0.0007670727791264653,
113        0.00048597707063890994,
114        0.00024241428764071316,
115        0.0003992895071860403,
116        0.0003530288813635707
117      ]
118    },
119    {
120      "cfg": {
121        "lr": 0.001,
122        "cycle_weight": 0.03
123      },
124      "mean": 0.0023312663834076375,
125      "std": 0.0007589181455697688,
126      "per_seed": [
127        0.0011521552223712206,
128        0.002354319440200925,
129        0.0021086849737912416,
130        0.002465922152623534,
131        0.0014372181612998247,
132        0.0023133959621191025,
133        0.0035222345031797886,
134        0.0032962006516754627
135      ]
136    },
137    {
138      "cfg": {
139        "lr": 0.01,
140        "cycle_weight": 0.03
141      },
142      "mean": 0.00045470128861779813,
143      "std": 0.00014832940019044596,
144      "per_seed": [
145        0.0004892470315098763,
146        0.0003560753830242902,
147        0.0005445053684525192,
148        0.0007670727791264653,
149        0.00048597707063890994,
150        0.00024241428764071316,
151        0.0003992895071860403,
152        0.0003530288813635707
153      ]
154    }
155  ],
156  "notes": "Matched dynamics task; baseline canonical train_model, idea same GRU backbone with differentiable factor-state regularizer."
157}