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      "gamma": 1.0
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "gamma": 0.4
 17        },
 18        "mean": 0.001971690042410046
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.001,
 23          "gamma": 1.0
 24        },
 25        "mean": 0.0013888165121898055
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.003,
 30          "gamma": 0.4
 31        },
 32        "mean": 0.0010129157599294558
 33      },
 34      {
 35        "cfg": {
 36          "lr": 0.003,
 37          "gamma": 1.0
 38        },
 39        "mean": 0.0007739054271951318
 40      },
 41      {
 42        "cfg": {
 43          "lr": 0.006,
 44          "gamma": 0.4
 45        },
 46        "mean": 0.0005326326572685502
 47      },
 48      {
 49        "cfg": {
 50          "lr": 0.006,
 51          "gamma": 1.0
 52        },
 53        "mean": 0.0004311748270993121
 54      }
 55    ],
 56    "full": {
 57      "mean": 0.0004004990041721612,
 58      "std": 7.354865862466145e-05,
 59      "per_seed": [
 60        0.00032163842115551233,
 61        0.0005050268373452127,
 62        0.0003722184046637267,
 63        0.0005258156452327967,
 64        0.00031021691393107176,
 65        0.00036279112100601196,
 66        0.00040758162504062057,
 67        0.0003987030650023371
 68      ],
 69      "n": 8
 70    }
 71  },
 72  "idea": {
 73    "mean": 0.0005187966198718641,
 74    "std": 6.731488481596035e-05,
 75    "per_seed": [
 76      0.0004997514770366251,
 77      0.0006125298095867038,
 78      0.0005725970841012895,
 79      0.000530327029991895,
 80      0.00038920878432691097,
 81      0.0005293103749863803,
 82      0.0005672792322002351,
 83      0.00044936916674487293
 84    ],
 85    "n": 8
 86  },
 87  "comparison": {
 88    "delta_mean": 0.00011829761569970287,
 89    "idea_wins": 0,
 90    "n_pairs": 8,
 91    "per_seed_diffs": [
 92      0.00017811305588111281,
 93      0.00010750297224149108,
 94      0.00020037867943756282,
 95      4.511384759098291e-06,
 96      7.899187039583921e-05,
 97      0.00016651925398036838,
 98      0.00015969760715961456,
 99      5.066610174253583e-05
100    ],
101    "p_value": 0.0081,
102    "mde": 5.78763280347936e-05,
103    "mde_rel_pct": 14.451054167893634,
104    "verdict": "idea worse (significant)",
105    "system_worked": false
106  },
107  "mechanism_signature": {
108    "prediction": "trained fast/slow modal decay rates agree within 20%",
109    "predicted_rates": [
110      3.908752053976059,
111      0.3747453838586807
112    ],
113    "observed_rates": [
114      0.0,
115      0.0
116    ],
117    "relative_errors": [
118      1.0,
119      1.0
120    ],
121    "confirmed": false
122  },
123  "idea_sweep": [
124    {
125      "cfg": {
126        "lr": 0.001
127      },
128      "mean": 0.001689480894128792,
129      "per_seed": [
130        0.0016606806311756372,
131        0.0017868814757093787,
132        0.0018675795290619135,
133        0.0017374568851664662,
134        0.0013423769269138575,
135        0.0014587605837732553,
136        0.001253080670721829,
137        0.0024090304505079985
138      ]
139    },
140    {
141      "cfg": {
142        "lr": 0.003
143      },
144      "mean": 0.0008820996736176312,
145      "per_seed": [
146        0.00092177907936275,
147        0.0009301036479882896,
148        0.0009488135110586882,
149        0.0009710948215797544,
150        0.0007397678564302623,
151        0.0008799024508334696,
152        0.0008156044641509652,
153        0.0008497315575368702
154      ]
155    },
156    {
157      "cfg": {
158        "lr": 0.006
159      },
160      "mean": 0.0005187966198718641,
161      "per_seed": [
162        0.0004997514770366251,
163        0.0006125298095867038,
164        0.0005725970841012895,
165        0.000530327029991895,
166        0.00038920878432691097,
167        0.0005293103749863803,
168        0.0005672792322002351,
169        0.00044936916674487293
170      ]
171    }
172  ],
173  "protocol_note": "Matched dynamics task; baseline and idea share 64-dimensional dissipative recurrent state, input/output, optimizer, epochs, batch, and all learning-rate values. Only one versus two relaxation bands and explicit exchange differ."
174}