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": 0.009,
 10      "epochs": 20
 11    },
 12    "sweep": [
 13      {
 14        "config": {
 15          "lr": 0.001,
 16          "epochs": 20
 17        },
 18        "mean": 0.0014746972592547536,
 19        "per_seed": [
 20          0.0012720254017040133,
 21          0.0014296569861471653,
 22          0.0013667949242517352,
 23          0.0018303117249161005
 24        ]
 25      },
 26      {
 27        "config": {
 28          "lr": 0.003,
 29          "epochs": 20
 30        },
 31        "mean": 0.0009116442524828017,
 32        "per_seed": [
 33          0.0007490376010537148,
 34          0.000787030381616205,
 35          0.0009362144046463072,
 36          0.0011742946226149797
 37        ]
 38      },
 39      {
 40        "config": {
 41          "lr": 0.009,
 42          "epochs": 20
 43        },
 44        "mean": 0.00042038498941110447,
 45        "per_seed": [
 46          0.0002886170113924891,
 47          0.0003179473278578371,
 48          0.00030860063270665705,
 49          0.0007663749856874347
 50        ]
 51      }
 52    ],
 53    "full": {
 54      "per_seed": [
 55        0.0002886170113924891,
 56        0.0003179473278578371,
 57        0.00030860063270665705,
 58        0.0007663749856874347,
 59        0.0003769927134271711,
 60        0.00034989885170944035,
 61        0.0004854405706282705,
 62        0.0004627566668204963
 63      ],
 64      "mean": 0.0004195785950287245
 65    }
 66  },
 67  "idea": {
 68    "config": {
 69      "lr": 0.009,
 70      "epochs": 20,
 71      "comm_lambda": 0.05
 72    },
 73    "per_seed": [
 74      1.915471875690855e-05,
 75      5.7846078561851755e-05,
 76      4.466433892957866e-05,
 77      1.5033640011097305e-05,
 78      1.8127400835510343e-05,
 79      3.6396970244823024e-05,
 80      2.411130117252469e-05,
 81      2.4222023057518527e-05
 82    ],
 83    "mean": 2.9944558946226607e-05
 84  },
 85  "comparison": {
 86    "delta_mean": -0.0003896340360824979,
 87    "idea_wins": 8,
 88    "n_pairs": 8,
 89    "per_seed_diffs": [
 90      -0.0002694622926355805,
 91      -0.0002601012492959853,
 92      -0.0002639362937770784,
 93      -0.0007513413456763374,
 94      -0.00035886531259166077,
 95      -0.0003135018814646173,
 96      -0.0004613292694557458,
 97      -0.0004385346437629778
 98    ],
 99    "p_value": 0.0081,
100    "mde": 0.00013852579096433872,
101    "mde_rel_pct": 33.01545708137356,
102    "verdict": "idea better (significant)",
103    "system_worked": true
104  },
105  "track_justification": "Dynamics is structurally matched to switched latent stability and Lyapunov contraction.",
106  "mechanism_signature": {
107    "prediction": "commutator regularization lowers trained ordered-cycle noncommutativity and cycle gain",
108    "observed": {
109      "observed_commutator_mean": 0.0003028972596865158,
110      "observed_cycle_gain_mean": 0.9300689920783043,
111      "n_models": 8
112    },
113    "baseline_observed": {
114      "cycle_gain_mean": NaN,
115      "input_sensitivity_mean": 0.09274659305810928,
116      "note": "GRU has no explicit matrix-cycle generator"
117    },
118    "idea_input_sensitivity_mean": 0.07906531821936369,
119    "confirmed": false
120  },
121  "idea_sweep": [
122    {
123      "config": {
124        "lr": 0.009,
125        "epochs": 20,
126        "comm_lambda": 0.0
127      },
128      "mean": 2.995237787217775e-05,
129      "per_seed": [
130        1.9160790543537587e-05,
131        5.785106623079628e-05,
132        4.468455517780967e-05,
133        1.5057028576848097e-05,
134        1.8133563571609557e-05,
135        3.6374363844515756e-05,
136        2.4118378860293888e-05,
137        2.423927617201116e-05
138      ]
139    },
140    {
141      "config": {
142        "lr": 0.009,
143        "epochs": 20,
144        "comm_lambda": 0.01
145      },
146      "mean": 2.995091836055508e-05,
147      "per_seed": [
148        1.9159475414198823e-05,
149        5.7849942095344886e-05,
150        4.4682146835839376e-05,
151        1.5051737136673182e-05,
152        1.813237940950785e-05,
153        3.637870031525381e-05,
154        2.4117280190694146e-05,
155        2.423568548692856e-05
156      ]
157    },
158    {
159      "config": {
160        "lr": 0.009,
161        "epochs": 20,
162        "comm_lambda": 0.05
163      },
164      "mean": 2.9944558946226607e-05,
165      "per_seed": [
166        1.915471875690855e-05,
167        5.7846078561851755e-05,
168        4.466433892957866e-05,
169        1.5033640011097305e-05,
170        1.8127400835510343e-05,
171        3.6396970244823024e-05,
172        2.411130117252469e-05,
173        2.4222023057518527e-05
174      ]
175    }
176  ]
177}