1{
  2  "bench_version": 1,
  3  "track": "sequence",
  4  "model": "transformer_tiny",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "sweep": {
  9      "best_cfg": {
 10        "lr": 0.001,
 11        "weight_decay": 0.0
 12      },
 13      "sweep": [
 14        {
 15          "cfg": {
 16            "lr": 0.001,
 17            "weight_decay": 0.0
 18          },
 19          "mean": 0.4265454038977623
 20        },
 21        {
 22          "cfg": {
 23            "lr": 0.001,
 24            "weight_decay": 0.0001
 25          },
 26          "mean": 0.42682382464408875
 27        },
 28        {
 29          "cfg": {
 30            "lr": 0.003,
 31            "weight_decay": 0.0
 32          },
 33          "mean": 0.5947448164224625
 34        },
 35        {
 36          "cfg": {
 37            "lr": 0.003,
 38            "weight_decay": 0.0001
 39          },
 40          "mean": 0.5931270122528076
 41        },
 42        {
 43          "cfg": {
 44            "lr": 0.006,
 45            "weight_decay": 0.0
 46          },
 47          "mean": 0.777096152305603
 48        },
 49        {
 50          "cfg": {
 51            "lr": 0.006,
 52            "weight_decay": 0.0001
 53          },
 54          "mean": 0.7612117379903793
 55        }
 56      ],
 57      "full": {
 58        "mean": 0.4578530229628086,
 59        "std": 0.07225049830230441,
 60        "per_seed": [
 61          0.4550389349460602,
 62          0.39718925952911377,
 63          0.41401442885398865,
 64          0.4399389922618866,
 65          0.6075195670127869,
 66          0.43840858340263367,
 67          0.536102831363678,
 68          0.37461158633232117
 69        ],
 70        "n": 8
 71      }
 72    },
 73    "best_config": {
 74      "lr": 0.001,
 75      "weight_decay": 0.0
 76    },
 77    "full": {
 78      "per_seed": [
 79        0.4550389349460602,
 80        0.39718925952911377,
 81        0.41401442885398865,
 82        0.4399389922618866,
 83        0.6075195670127869,
 84        0.43840858340263367,
 85        0.536102831363678,
 86        0.37461158633232117
 87      ],
 88      "mean": 0.4578530229628086
 89    }
 90  },
 91  "idea": {
 92    "configs_tested": [
 93      {
 94        "lr": 0.001,
 95        "weight_decay": 0.0
 96      },
 97      {
 98        "lr": 0.003,
 99        "weight_decay": 0.0
100      },
101      {
102        "lr": 0.003,
103        "weight_decay": 0.0001
104      }
105    ],
106    "best_config": {
107      "lr": 0.001,
108      "weight_decay": 0.0
109    },
110    "per_seed": [
111      0.343902587890625,
112      0.3311694264411926,
113      0.45799392461776733,
114      0.3579498827457428,
115      0.4836745262145996,
116      0.8792790174484253,
117      0.5392436981201172,
118      0.7493586540222168
119    ],
120    "mean": 0.5178214646875858
121  },
122  "comparison": {
123    "delta_mean": 0.05996844172477722,
124    "idea_wins": 4,
125    "n_pairs": 8,
126    "per_seed_diffs": [
127      -0.11113634705543518,
128      -0.06601983308792114,
129      0.043979495763778687,
130      -0.0819891095161438,
131      -0.12384504079818726,
132      0.4408704340457916,
133      0.003140866756439209,
134      0.37474706768989563
135    ],
136    "p_value": 0.4928,
137    "mde": 0.1860985402118441,
138    "mde_rel_pct": 40.64591274457106,
139    "verdict": "no significant win",
140    "system_worked": false
141  },
142  "protocol": {
143    "epochs": 12,
144    "n_train": 400,
145    "n_test": 200,
146    "alpha": 1.0,
147    "grid_union": [
148      {
149        "lr": 0.001,
150        "weight_decay": 0.0
151      },
152      {
153        "lr": 0.001,
154        "weight_decay": 0.0001
155      },
156      {
157        "lr": 0.003,
158        "weight_decay": 0.0
159      },
160      {
161        "lr": 0.003,
162        "weight_decay": 0.0001
163      },
164      {
165        "lr": 0.006,
166        "weight_decay": 0.0
167      },
168      {
169        "lr": 0.006,
170        "weight_decay": 0.0001
171      }
172    ]
173  },
174  "mechanism_signature": {
175    "prediction": "whitening reduces temporal autocorrelation and f_y approximately equals f_hat^-1 f",
176    "observed_lag_energy_reduction_mean": -0.08987211176084602,
177    "observed_predicted_flatten_ratio_mean": 0.9619392576036487,
178    "predicted_lag_energy_reduction": "positive",
179    "predicted_flatten_ratio": 1.0,
180    "confirmed": false
181  }
182}