Spectrally safeguarded DFP / bench_report.json

✓✓ Beats tuned baseline

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      "epochs": 8
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "epochs": 8
 17        },
 18        "mean": 217.77801513671875
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "epochs": 8
 24        },
 25        "mean": 149.99735641479492
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.01,
 30          "epochs": 8
 31        },
 32        "mean": 26.118374347686768
 33      }
 34    ],
 35    "full": {
 36      "mean": 35.374396085739136,
 37      "std": 12.759078946030915,
 38      "per_seed": [
 39        29.38343620300293,
 40        32.629642486572266,
 41        22.46150779724121,
 42        19.998910903930664,
 43        34.86793899536133,
 44        63.3486328125,
 45        36.457733154296875,
 46        43.84736633300781
 47      ],
 48      "n": 8
 49    }
 50  },
 51  "idea": {
 52    "mean": 13.662639737129211,
 53    "std": 1.2255122413284667,
 54    "per_seed": [
 55      13.658586502075195,
 56      10.993560791015625,
 57      14.850048065185547,
 58      14.97375202178955,
 59      12.930068016052246,
 60      13.51315689086914,
 61      13.623165130615234,
 62      14.758780479431152
 63    ],
 64    "n": 8,
 65    "sweep": [
 66      {
 67        "mean": 16.703367471694946,
 68        "std": 1.902843853945096,
 69        "per_seed": [
 70          16.672517776489258,
 71          12.807530403137207,
 72          17.32662010192871,
 73          16.859230041503906,
 74          15.626265525817871,
 75          18.918928146362305,
 76          16.08186149597168,
 77          19.333986282348633
 78        ],
 79        "n": 8,
 80        "cfg": {
 81          "lr": 0.01
 82        }
 83      },
 84      {
 85        "mean": 31.656866788864136,
 86        "std": 7.626313276119716,
 87        "per_seed": [
 88          25.898263931274414,
 89          24.708200454711914,
 90          19.921588897705078,
 91          45.958133697509766,
 92          34.698883056640625,
 93          37.070926666259766,
 94          33.396942138671875,
 95          31.60199546813965
 96        ],
 97        "n": 8,
 98        "cfg": {
 99          "lr": 0.005
100        }
101      },
102      {
103        "mean": 13.662639737129211,
104        "std": 1.2255122413284667,
105        "per_seed": [
106          13.658586502075195,
107          10.993560791015625,
108          14.850048065185547,
109          14.97375202178955,
110          12.930068016052246,
111          13.51315689086914,
112          13.623165130615234,
113          14.758780479431152
114        ],
115        "n": 8,
116        "cfg": {
117          "lr": 0.02
118        }
119      }
120    ],
121    "best_cfg": {
122      "lr": 0.02
123    }
124  },
125  "comparison": {
126    "delta_mean": -21.711756348609924,
127    "idea_wins": 8,
128    "n_pairs": 8,
129    "per_seed_diffs": [
130      -15.724849700927734,
131      -21.63608169555664,
132      -7.611459732055664,
133      -5.025158882141113,
134      -21.937870979309082,
135      -49.83547592163086,
136      -22.83456802368164,
137      -29.08858585357666
138    ],
139    "p_value": 0.0081,
140    "mde": 11.653916222940111,
141    "mde_rel_pct": 32.94449520691119,
142    "verdict": "idea better (significant)",
143    "system_worked": true
144  },
145  "mechanism_signature": {
146    "prediction": "spectral floor and rotation safeguard bound inverse-Hessian spectrum",
147    "predicted_floor": 0.0001,
148    "observed_min_eig_min": 0.00015548363137000883,
149    "observed_max_rotation": 0.9999999997777941,
150    "safeguards_total": 374,
151    "confirmed": true
152  },
153  "math_check": {
154    "secant_error": 3.1249111476034713e-15,
155    "floor_min_eig": 0.001,
156    "secant_pass": true,
157    "floor_pass": true
158  }
159}