1{
  2  "bench_version": 1,
  3  "track": "long_1d_convolution",
  4  "model": "cnn_small",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.01,
 10      "m": 8
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "m": 8
 17        },
 18        "mean": 0.4312499910593033
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "m": 8
 24        },
 25        "mean": 0.4387499913573265
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.01,
 30          "m": 8
 31        },
 32        "mean": 0.35874998569488525
 33      }
 34    ],
 35    "full": {
 36      "mean": 0.35249999165534973,
 37      "std": 0.02772634371136278,
 38      "per_seed": [
 39        0.41499999165534973,
 40        0.35999998450279236,
 41        0.32999998331069946,
 42        0.32999998331069946,
 43        0.3449999988079071,
 44        0.3700000047683716,
 45        0.3449999988079071,
 46        0.32499998807907104
 47      ],
 48      "n": 8
 49    }
 50  },
 51  "idea": {
 52    "mean": 0.48499998450279236,
 53    "std": 0.03381936953316897,
 54    "per_seed": [
 55      0.429999977350235,
 56      0.4399999976158142,
 57      0.5299999713897705,
 58      0.5149999856948853,
 59      0.4899999797344208,
 60      0.4699999988079071,
 61      0.5149999856948853,
 62      0.4899999797344208
 63    ],
 64    "n": 8
 65  },
 66  "comparison": {
 67    "delta_mean": 0.13249999284744263,
 68    "idea_wins": 0,
 69    "n_pairs": 8,
 70    "per_seed_diffs": [
 71      0.014999985694885254,
 72      0.08000001311302185,
 73      0.19999998807907104,
 74      0.1850000023841858,
 75      0.14499998092651367,
 76      0.09999999403953552,
 77      0.16999998688697815,
 78      0.16499999165534973
 79    ],
 80    "p_value": 0.0081,
 81    "mde": 0.05245633272675812,
 82    "mde_rel_pct": 14.881229494622602,
 83    "verdict": "idea worse (significant)",
 84    "system_worked": false
 85  },
 86  "mechanism_signature": {
 87    "mechanism_signature": {
 88      "prediction": "implicit convolution equals explicit convolution; tile storage ratio M/(ceil(L/m)*m)",
 89      "predicted_relative_operator_error": 0.0,
 90      "observed_relative_operator_error": 1.0825973749160767,
 91      "predicted_storage_ratio": 2.0,
 92      "trained_system_output_gap": 1.0547434091567993,
 93      "confirmed": false
 94    },
 95    "custom_track": {
 96      "name": "long_1d_convolution",
 97      "file": "stage2_bench.py",
 98      "domain": "cnn"
 99    }
100  },
101  "idea_sweep": [
102    {
103      "cfg": {
104        "lr": 0.001,
105        "m": 8
106      },
107      "mean": 0.48499999195337296
108    },
109    {
110      "cfg": {
111        "lr": 0.003,
112        "m": 8
113      },
114      "mean": 0.48624999076128006
115    },
116    {
117      "cfg": {
118        "lr": 0.01,
119        "m": 8
120      },
121      "mean": 0.47874998301267624
122    }
123  ],
124  "selection_note": "Baseline and idea share lr union [0.001, 0.003, 0.01]; idea m fixed a priori at 8."
125}