1{
  2  "bench_version": 1,
  3  "track": "tabular",
  4  "model": "mlp_tiny",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.003,
 10      "weight_decay": 0.0
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "weight_decay": 0.0
 17        },
 18        "mean": 188.35083389282227
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "weight_decay": 0.0
 24        },
 25        "mean": 19.947585344314575
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.006,
 30          "weight_decay": 0.0
 31        },
 32        "mean": 22.523260593414307
 33      }
 34    ],
 35    "full": {
 36      "mean": 21.810271620750427,
 37      "std": 6.031322024802775,
 38      "per_seed": [
 39        17.455322265625,
 40        15.50521183013916,
 41        29.37735366821289,
 42        17.45245361328125,
 43        33.27134323120117,
 44        16.96589469909668,
 45        22.887645721435547,
 46        21.56694793701172
 47      ],
 48      "n": 8
 49    }
 50  },
 51  "idea": {
 52    "mean": 21.595937252044678,
 53    "std": 5.868687330662625,
 54    "per_seed": [
 55      17.42387580871582,
 56      15.323787689208984,
 57      28.447860717773438,
 58      17.114221572875977,
 59      32.90999221801758,
 60      16.928159713745117,
 61      22.993961334228516,
 62      21.625638961791992
 63    ],
 64    "n": 8
 65  },
 66  "comparison": {
 67    "delta_mean": -0.2143343687057495,
 68    "idea_wins": 6,
 69    "n_pairs": 8,
 70    "per_seed_diffs": [
 71      -0.03144645690917969,
 72      -0.18142414093017578,
 73      -0.9294929504394531,
 74      -0.33823204040527344,
 75      -0.36135101318359375,
 76      -0.0377349853515625,
 77      0.10631561279296875,
 78      0.05869102478027344
 79    ],
 80    "p_value": 0.09385,
 81    "mde": 0.2811647999151475,
 82    "mde_rel_pct": 1.2891393780151073,
 83    "verdict": "no measurable effect",
 84    "system_worked": false
 85  },
 86  "mechanism_signature": {
 87    "prediction": {
 88      "feature_dim": 64,
 89      "observed_rank_last": 64,
 90      "predicted_identity_max_eigen_mean": 1.8750072717666626,
 91      "observed_QtQ_error_max": 1.0,
 92      "prediction": "MGS memory predicts identity Gramian after d independent directions",
 93      "confirmed": false
 94    },
 95    "idea_configs": [
 96      {
 97        "cfg": {
 98          "lr": 0.003,
 99          "lambda": 0.02
100        },
101        "result": {
102          "mean": 21.754212260246277,
103          "std": 5.9932442203746445,
104          "per_seed": [
105            17.445890426635742,
106            15.479540824890137,
107            29.154930114746094,
108            17.36090087890625,
109            33.19727325439453,
110            16.9453067779541,
111            22.87087631225586,
112            21.5789794921875
113          ],
114          "n": 8
115        }
116      },
117      {
118        "cfg": {
119          "lr": 0.003,
120          "lambda": 0.05
121        },
122        "result": {
123          "mean": 21.67228603363037,
124          "std": 5.945718685399587,
125          "per_seed": [
126            17.41834259033203,
127            15.36572265625,
128            28.8680419921875,
129            17.256317138671875,
130            33.06776428222656,
131            16.93799591064453,
132            22.85132598876953,
133            21.612777709960938
134          ],
135          "n": 8
136        }
137      },
138      {
139        "cfg": {
140          "lr": 0.003,
141          "lambda": 0.1
142        },
143        "result": {
144          "mean": 21.595937252044678,
145          "std": 5.868687330662625,
146          "per_seed": [
147            17.42387580871582,
148            15.323787689208984,
149            28.447860717773438,
150            17.114221572875977,
151            32.90999221801758,
152            16.928159713745117,
153            22.993961334228516,
154            21.625638961791992
155          ],
156          "n": 8
157        }
158      }
159    ],
160    "confirmed": false
161  },
162  "track_justification": "Tabular is the prescribed optimizer track; Friedman regression with mlp_tiny has a trainable final linear head and fixed feature dimension, matching the memory correction."
163}