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    },
 11    "sweep": [
 12      {
 13        "cfg": {
 14          "lr": 0.001
 15        },
 16        "mean": 0.010650048730894923
 17      },
 18      {
 19        "cfg": {
 20          "lr": 0.003
 21        },
 22        "mean": 0.0021922332525718957
 23      },
 24      {
 25        "cfg": {
 26          "lr": 0.01
 27        },
 28        "mean": 0.0001847558050940279
 29      }
 30    ],
 31    "full": {
 32      "mean": 0.00020586385835485999,
 33      "std": 5.912498369648286e-05,
 34      "per_seed": [
 35        0.00017305620713159442,
 36        0.00018997277948074043,
 37        0.00013568498252425343,
 38        0.0002403092512395233,
 39        0.00026725101633928716,
 40        0.00018888202612288296,
 41        0.0003143867361359298,
 42        0.00013736786786466837
 43      ],
 44      "n": 8
 45    },
 46    "idea_union_sweep": [
 47      {
 48        "cfg": {
 49          "lr": 0.01,
 50          "K": 4
 51        },
 52        "mean": 0.000600305080297403
 53      },
 54      {
 55        "cfg": {
 56          "lr": 0.001,
 57          "K": 4
 58        },
 59        "mean": 0.0048645936767570674
 60      },
 61      {
 62        "cfg": {
 63          "lr": 0.01,
 64          "K": 4
 65        },
 66        "mean": 0.000600305080297403
 67      }
 68    ]
 69  },
 70  "idea": {
 71    "mean": 0.0005137363186804578,
 72    "std": 0.00021520799203052408,
 73    "per_seed": [
 74      0.0005184079636819661,
 75      0.00034054333809763193,
 76      0.000662883510813117,
 77      0.0008793855085968971,
 78      0.00037677702493965626,
 79      0.00013013248099014163,
 80      0.0005593107198365033,
 81      0.0006424500024877489
 82    ],
 83    "n": 8
 84  },
 85  "comparison": {
 86    "delta_mean": 0.0003078724603255978,
 87    "idea_wins": 1,
 88    "n_pairs": 8,
 89    "per_seed_diffs": [
 90      0.00034535175655037165,
 91      0.0001505705586168915,
 92      0.0005271985282888636,
 93      0.0006390762573573738,
 94      0.0001095260086003691,
 95      -5.874954513274133e-05,
 96      0.00024492398370057344,
 97      0.0005050821346230805
 98    ],
 99    "p_value": 0.0163,
100    "mde": 0.00020001775934380757,
101    "mde_rel_pct": 97.16021109398662,
102    "verdict": "idea worse (significant)",
103    "system_worked": false
104  },
105  "mechanism_signature": {
106    "prediction": "analytic triangular inverse and logdet remain exact after training",
107    "reconstruction_max_abs": 1.1920928955078125e-07,
108    "observed_mean_abs_logdet": 0.786198616027832,
109    "observed_jacobian_logdet_abs_error": 1.1920928955078125e-07,
110    "jacobian_sign": 1.0,
111    "confirmed": true
112  },
113  "protocol_notes": {
114    "epochs": 12,
115    "n_train": 800,
116    "n_test": 300,
117    "idea_configs": [
118      {
119        "lr": 0.01,
120        "K": 4
121      },
122      {
123        "lr": 0.001,
124        "K": 4
125      },
126      {
127        "lr": 0.01,
128        "K": 4
129      }
130    ],
131    "structural_match": "controlled pendulum multi-step rollout; flow is trained end-to-end with shared GRU encoder"
132  }
133}