1{
  2  "bench_version": 1,
  3  "track": "complex_array_channel",
  4  "model": "mlp_tiny",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.01,
 10      "k": 2
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "k": 2
 17        },
 18        "mean": 0.007402819930575788
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "k": 2
 24        },
 25        "mean": 0.0065840884344652295
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.01,
 30          "k": 2
 31        },
 32        "mean": 0.006486733560450375
 33      }
 34    ],
 35    "full": {
 36      "mean": 0.006718522927258164,
 37      "std": 0.0007129592821712252,
 38      "per_seed": [
 39        0.006483251228928566,
 40        0.00544784078374505,
 41        0.007208648603409529,
 42        0.006807193625718355,
 43        0.007167691830545664,
 44        0.007836345583200455,
 45        0.005902399308979511,
 46        0.006894812453538179
 47      ],
 48      "n": 8
 49    }
 50  },
 51  "idea": {
 52    "mean": 0.007619430776685476,
 53    "std": 0.0008275326682381549,
 54    "per_seed": [
 55      0.0074270921759307384,
 56      0.006459011230617762,
 57      0.008539552800357342,
 58      0.007304489612579346,
 59      0.007882530800998211,
 60      0.00901730451732874,
 61      0.006567466538399458,
 62      0.007757998537272215
 63    ],
 64    "n": 8
 65  },
 66  "comparison": {
 67    "delta_mean": 0.0009009078494273126,
 68    "idea_wins": 0,
 69    "n_pairs": 8,
 70    "per_seed_diffs": [
 71      0.0009438409470021725,
 72      0.0010111704468727112,
 73      0.001330904196947813,
 74      0.0004972959868609905,
 75      0.0007148389704525471,
 76      0.0011809589341282845,
 77      0.0006650672294199467,
 78      0.0008631860837340355
 79    ],
 80    "p_value": 0.0081,
 81    "mde": 0.0002304888187870338,
 82    "mde_rel_pct": 3.43064720151363,
 83    "verdict": "idea worse (significant)",
 84    "system_worked": false
 85  },
 86  "mechanism_signature": {
 87    "prediction": "Taylor steering error scales quadratically in offset",
 88    "delta": 0.001,
 89    "predicted_error_order": 2.0,
 90    "observed_error_over_delta_squared": 34.418742870911956,
 91    "trained_model_output_rms": 0.011084591038525105,
 92    "confirmed": true
 93  },
 94  "idea_sweep": [
 95    {
 96      "cfg": {
 97        "lr": 0.001,
 98        "k": 2
 99      },
100      "mean": 0.007869121560361236,
101      "std": 0.0008392636492011609,
102      "per_seed": [
103        0.007451348938047886,
104        0.006571656093001366,
105        0.008513221517205238,
106        0.007863671518862247,
107        0.008256588131189346,
108        0.009301764890551567,
109        0.006832886021584272,
110        0.008161835372447968
111      ],
112      "n": 8
113    },
114    {
115      "cfg": {
116        "lr": 0.003,
117        "k": 2
118      },
119      "mean": 0.007681989867705852,
120      "std": 0.0008145156626225875,
121      "per_seed": [
122        0.00713324686512351,
123        0.006568302866071463,
124        0.008564387448132038,
125        0.0075448122806847095,
126        0.008067510090768337,
127        0.0090027442201972,
128        0.0066692763939499855,
129        0.00790563877671957
130      ],
131      "n": 8
132    },
133    {
134      "cfg": {
135        "lr": 0.01,
136        "k": 2
137      },
138      "mean": 0.007619430776685476,
139      "std": 0.0008275326682381549,
140      "per_seed": [
141        0.0074270921759307384,
142        0.006459011230617762,
143        0.008539552800357342,
144        0.007304489612579346,
145        0.007882530800998211,
146        0.00901730451732874,
147        0.006567466538399458,
148        0.007757998537272215
149      ],
150      "n": 8
151    }
152  ],
153  "custom_track": {
154    "name": "complex_array_channel",
155    "file": "array_track.py",
156    "domain": "array-valued complex low-rank tensors"
157  }
158}