FDT-Calibrated Rotational Optimizer / bench_report.json

Mechanism confirmed, baseline not beaten

Raw ⬇ ZIP
  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": 0.01,
 10      "weight_decay": 0.0001
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "weight_decay": 0.0
 17        },
 18        "mean": 130.23694229125977
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.001,
 23          "weight_decay": 0.0001
 24        },
 25        "mean": 130.24530410766602
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.003,
 30          "weight_decay": 0.0
 31        },
 32        "mean": 17.47633147239685
 33      },
 34      {
 35        "cfg": {
 36          "lr": 0.003,
 37          "weight_decay": 0.0001
 38        },
 39        "mean": 17.535193920135498
 40      },
 41      {
 42        "cfg": {
 43          "lr": 0.01,
 44          "weight_decay": 0.0
 45        },
 46        "mean": 11.130715131759644
 47      },
 48      {
 49        "cfg": {
 50          "lr": 0.01,
 51          "weight_decay": 0.0001
 52        },
 53        "mean": 11.037371158599854
 54      }
 55    ],
 56    "full": {
 57      "mean": 11.380033016204834,
 58      "std": 1.038192373606595,
 59      "per_seed": [
 60        10.492472648620605,
 61        12.06095027923584,
 62        10.988505363464355,
 63        10.607556343078613,
 64        9.646132469177246,
 65        11.897088050842285,
 66        12.67528247833252,
 67        12.672276496887207
 68      ],
 69      "n": 8
 70    }
 71  },
 72  "idea": {
 73    "mean": 19.045888423919678,
 74    "std": 11.022546935963494,
 75    "per_seed": [
 76      10.960366249084473,
 77      18.753145217895508,
 78      37.41089630126953,
 79      12.765547752380371,
 80      9.869215965270996,
 81      14.048157691955566,
 82      10.755529403686523,
 83      37.80424880981445
 84    ],
 85    "n": 8
 86  },
 87  "comparison": {
 88    "delta_mean": 7.665855407714844,
 89    "idea_wins": 1,
 90    "n_pairs": 8,
 91    "per_seed_diffs": [
 92      0.4678936004638672,
 93      6.692194938659668,
 94      26.422390937805176,
 95      2.157991409301758,
 96      0.22308349609375,
 97      2.1510696411132812,
 98      -1.919753074645996,
 99      25.131972312927246
100    ],
101    "p_value": 0.0394,
102    "mde": 9.573659179917387,
103    "mde_rel_pct": 84.12681374724288,
104    "verdict": "idea worse (significant)",
105    "system_worked": false
106  },
107  "mechanism_signature": {
108    "prediction": "skew component is orthogonal to current gradient and increases complex/rotational response while stability requires rho<1",
109    "observed_jacobian_rayleigh": -1.708630919456482,
110    "observed_euler_rayleigh_factor": 1.0051258927583695,
111    "observed_rotation_gradient_cosine": -5.9046900702242056e-09,
112    "observed_skew_to_gradient_norm": 0.5999951958656311,
113    "confirmed": true
114  },
115  "idea_sweep": [
116    {
117      "cfg": {
118        "lr": 0.01,
119        "alpha": 0.0
120      },
121      "mean": 29.906624674797058,
122      "std": 26.379266590239755,
123      "per_seed": [
124        39.70406723022461,
125        17.099863052368164,
126        10.755126953125,
127        14.188899040222168,
128        11.864416122436523,
129        10.44906234741211,
130        43.637142181396484,
131        91.5544204711914
132      ]
133    },
134    {
135      "cfg": {
136        "lr": 0.01,
137        "alpha": 0.3
138      },
139      "mean": 21.96018660068512,
140      "std": 21.920466510288684,
141      "per_seed": [
142        11.32851791381836,
143        15.652362823486328,
144        12.332928657531738,
145        13.191065788269043,
146        18.957841873168945,
147        13.081674575805664,
148        11.51638126373291,
149        79.62071990966797
150      ]
151    },
152    {
153      "cfg": {
154        "lr": 0.01,
155        "alpha": 0.6
156      },
157      "mean": 19.045888423919678,
158      "std": 11.022546935963494,
159      "per_seed": [
160        10.960366249084473,
161        18.753145217895508,
162        37.41089630126953,
163        12.765547752380371,
164        9.869215965270996,
165        14.048157691955566,
166        10.755529403686523,
167        37.80424880981445
168      ]
169    }
170  ],
171  "budget": {
172    "epochs": 12,
173    "batch": 128,
174    "n_train": 400,
175    "n_test": 400,
176    "seconds": 17.008500576019287
177  },
178  "custom_track": null
179}