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      "epochs": 8
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "epochs": 8
 17        },
 18        "mean": 0.0228443113155663
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "epochs": 8
 24        },
 25        "mean": 0.006279829074628651
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.01,
 30          "epochs": 8
 31        },
 32        "mean": 0.005194205295993015
 33      }
 34    ],
 35    "full": {
 36      "mean": 0.004214190877974033,
 37      "std": 0.0026682777355217157,
 38      "per_seed": [
 39        0.002182353986427188,
 40        0.008899358101189137,
 41        0.0017826670082286,
 42        0.007912442088127136,
 43        0.003060318995267153,
 44        0.0012072835816070437,
 45        0.0036685036029666662,
 46        0.005000599659979343
 47      ],
 48      "n": 8
 49    }
 50  },
 51  "idea": {
 52    "mean": 0.004029877163702622,
 53    "std": 0.0015034226018154422,
 54    "per_seed": [
 55      0.0031153184827417135,
 56      0.0026313886046409607,
 57      0.006552070379257202,
 58      0.0028610709123313427,
 59      0.005238592624664307,
 60      0.002022449392825365,
 61      0.004396332893520594,
 62      0.005421794019639492
 63    ],
 64    "n": 8,
 65    "selected_cfg": {
 66      "lr": 0.01,
 67      "epochs": 8
 68    }
 69  },
 70  "comparison": {
 71    "delta_mean": -0.0001843137142714113,
 72    "idea_wins": 2,
 73    "n_pairs": 8,
 74    "per_seed_diffs": [
 75      0.0009329644963145256,
 76      -0.006267969496548176,
 77      0.004769403371028602,
 78      -0.0050513711757957935,
 79      0.002178273629397154,
 80      0.0008151658112183213,
 81      0.0007278292905539274,
 82      0.0004211943596601486
 83    ],
 84    "p_value": 0.9399,
 85    "mde": 0.003068421703670571,
 86    "mde_rel_pct": 72.81164504693035,
 87    "verdict": "no significant win",
 88    "system_worked": false
 89  },
 90  "mechanism_signature": {
 91    "prediction": "recursive predicted variance should calibrate observed squared error near 1 and track heteroscedasticity",
 92    "predicted_vs_observed": {
 93      "mean_q_recursive": 0.8105446696281433,
 94      "mean_observed_squared_error": 0.005421794019639492,
 95      "calibration_ratio_mean": 0.005157590081528752,
 96      "calibration_ratio_per_seed": [
 97        0.003714949226948883,
 98        0.0034924976914672296,
 99        0.00926556392303059,
100        0.003918822999611737,
101        0.006340776802686519,
102        0.002460186822311339,
103        0.005378848268044676,
104        0.00668907491812904
105      ],
106      "q_error_correlation_mean": -0.09872341295990021
107    },
108    "confirmed": false
109  },
110  "protocol_notes": {
111    "structural_match": "multi-step stochastic actuated pendulum rollout; recursive uncertainty applies to transition branches",
112    "baseline": "standard MSE training",
113    "idea": "mean/variance RNN with detached recursive child q plus between-child variance",
114    "n_train": 400,
115    "n_test": 200,
116    "branch_count": 4,
117    "grid_union": [
118      {
119        "lr": 0.001,
120        "epochs": 8
121      },
122      {
123        "lr": 0.003,
124        "epochs": 8
125      },
126      {
127        "lr": 0.01,
128        "epochs": 8
129      }
130    ]
131  }
132}