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": 0.006,
 10      "epochs": 18,
 11      "weight_decay": 0.0
 12    },
 13    "sweep": [
 14      {
 15        "cfg": {
 16          "lr": 0.0015,
 17          "epochs": 18,
 18          "weight_decay": 0.0
 19        },
 20        "mean": 0.009193785954266787
 21      },
 22      {
 23        "cfg": {
 24          "lr": 0.003,
 25          "epochs": 18,
 26          "weight_decay": 0.0
 27        },
 28        "mean": 0.0050034308223985136
 29      },
 30      {
 31        "cfg": {
 32          "lr": 0.006,
 33          "epochs": 18,
 34          "weight_decay": 0.0
 35        },
 36        "mean": 0.002262153459014371
 37      }
 38    ],
 39    "full": {
 40      "mean": 0.001892383472295478,
 41      "std": 0.0005824768770598943,
 42      "per_seed": [
 43        0.0016680180560797453,
 44        0.0018032685620710254,
 45        0.002757444279268384,
 46        0.0028198829386383295,
 47        0.0014833298046141863,
 48        0.0013202878180891275,
 49        0.0011803944362327456,
 50        0.0021064418833702803
 51      ],
 52      "n": 8
 53    }
 54  },
 55  "idea": {
 56    "mean": 0.0024836512457113713,
 57    "std": 0.0007913986756607849,
 58    "per_seed": [
 59      0.0018633758882060647,
 60      0.002772978274151683,
 61      0.003546620486304164,
 62      0.003868584055453539,
 63      0.00169182859826833,
 64      0.001610392238944769,
 65      0.002222771057859063,
 66      0.002292659366503358
 67    ],
 68    "n": 8,
 69    "best_cfg": {
 70      "lr": 0.006,
 71      "epochs": 18,
 72      "weight_decay": 0.0,
 73      "q": 0.25,
 74      "noise": 0.02
 75    },
 76    "sweep": [
 77      {
 78        "cfg": {
 79          "lr": 0.006,
 80          "epochs": 18,
 81          "weight_decay": 0.0,
 82          "q": 0.25,
 83          "noise": 0.02
 84        },
 85        "mean": 0.0024836512457113713
 86      },
 87      {
 88        "cfg": {
 89          "lr": 0.006,
 90          "epochs": 18,
 91          "weight_decay": 0.0,
 92          "q": 0.5,
 93          "noise": 0.02
 94        },
 95        "mean": 0.0025042231136467308
 96      },
 97      {
 98        "cfg": {
 99          "lr": 0.006,
100          "epochs": 18,
101          "weight_decay": 0.0,
102          "q": 0.75,
103          "noise": 0.02
104        },
105        "mean": 0.0025412623363081366
106      }
107    ],
108    "selected_cfg": {
109      "lr": 0.006,
110      "epochs": 18,
111      "weight_decay": 0.0,
112      "q": 0.25,
113      "noise": 0.02
114    }
115  },
116  "comparison": {
117    "delta_mean": 0.0005912677734158933,
118    "idea_wins": 0,
119    "n_pairs": 8,
120    "per_seed_diffs": [
121      0.0001953578321263194,
122      0.0009697097120806575,
123      0.00078917620703578,
124      0.0010487011168152094,
125      0.0002084987936541438,
126      0.00029010442085564137,
127      0.0010423766216263175,
128      0.00018621748313307762
129    ],
130    "p_value": 0.0081,
131    "mde": 0.000339375496055173,
132    "mde_rel_pct": 17.933759252478964,
133    "verdict": "idea worse (significant)",
134    "system_worked": false
135  },
136  "mechanism_signature": {
137    "prediction": "dissipative chiral latent dynamics should remain stable and reduce local perturbation response",
138    "trained_model_observation": {
139      "baseline": {
140        "perturbation": 0.01,
141        "response_rms": 8.569472265662625e-05,
142        "output_rms": 0.1667313128709793
143      },
144      "idea": {
145        "perturbation": 0.01,
146        "response_rms": 8.628764771856368e-05,
147        "output_rms": 0.16530634462833405
148      }
149    },
150    "predicted_response_ratio": "<= 1.0",
151    "observed_response_ratio": 1.0069190264772103,
152    "confirmed": false
153  },
154  "structural_match": "dynamics: stability/control and multi-step pendulum rollout"
155}