Risk-Fitted Shrinkage Gate / bench_report.json

Failed on benchmark

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": 0.006,
 10      "weight_decay": 0.0001
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.0015,
 16          "weight_decay": 0.0
 17        },
 18        "mean": 16.81662940979004
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.0015,
 23          "weight_decay": 0.0001
 24        },
 25        "mean": 16.790626525878906
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.003,
 30          "weight_decay": 0.0
 31        },
 32        "mean": 13.860755920410156
 33      },
 34      {
 35        "cfg": {
 36          "lr": 0.003,
 37          "weight_decay": 0.0001
 38        },
 39        "mean": 13.839483976364136
 40      },
 41      {
 42        "cfg": {
 43          "lr": 0.006,
 44          "weight_decay": 0.0
 45        },
 46        "mean": 10.442761182785034
 47      },
 48      {
 49        "cfg": {
 50          "lr": 0.006,
 51          "weight_decay": 0.0001
 52        },
 53        "mean": 10.238675355911255
 54      }
 55    ],
 56    "full": {
 57      "mean": 10.678602695465088,
 58      "std": 1.2286671921700971,
 59      "per_seed": [
 60        10.056567192077637,
 61        11.58691120147705,
 62        9.562552452087402,
 63        9.74867057800293,
 64        8.789999961853027,
 65        11.738417625427246,
 66        11.37514591217041,
 67        12.570556640625
 68      ],
 69      "n": 8
 70    }
 71  },
 72  "idea": {
 73    "mean": 10.288059711456299,
 74    "std": 0.9213815344241648,
 75    "per_seed": [
 76      10.669325828552246,
 77      10.632473945617676,
 78      10.393888473510742,
 79      9.149672508239746,
 80      8.439580917358398,
 81      11.03823184967041,
 82      10.63460636138916,
 83      11.346697807312012
 84    ],
 85    "n": 8
 86  },
 87  "comparison": {
 88    "delta_mean": -0.39054298400878906,
 89    "idea_wins": 6,
 90    "n_pairs": 8,
 91    "per_seed_diffs": [
 92      0.6127586364746094,
 93      -0.954437255859375,
 94      0.8313360214233398,
 95      -0.5989980697631836,
 96      -0.3504190444946289,
 97      -0.7001857757568359,
 98      -0.74053955078125,
 99      -1.2238588333129883
100    ],
101    "p_value": 0.1779,
102    "mde": 0.6140738235288303,
103    "mde_rel_pct": 5.750507262430606,
104    "verdict": "no significant win",
105    "system_worked": false
106  },
107  "mechanism_signature": {
108    "quantity": "active hidden fraction after trained gate",
109    "predicted": "adaptive shrinkage should reduce active fraction",
110    "observed_mean": 0.999560534954071,
111    "observed_per_seed": [
112      0.999804675579071,
113      0.9994531273841858,
114      0.9996093511581421,
115      0.9993749856948853
116    ],
117    "confirmed": false
118  },
119  "math_check": {
120    "analytic_divergence": 3.467686653137207,
121    "finite_difference_divergence": 3.466308116912842,
122    "abs_error": 0.0013785362243652344,
123    "confirmed": true
124  },
125  "idea_sweep": [
126    {
127      "cfg": {
128        "lr": 0.006,
129        "sure_weight": 0.01
130      },
131      "mean": 10.211340188980103
132    },
133    {
134      "cfg": {
135        "lr": 0.006,
136        "sure_weight": 0.02
137      },
138      "mean": 10.212185859680176
139    },
140    {
141      "cfg": {
142        "lr": 0.006,
143        "sure_weight": 0.04
144      },
145      "mean": 10.213889598846436
146    }
147  ]
148}