LP-Embedded Input-Convex MLP / bench_report.json

Mechanism confirmed, baseline not beaten

Raw ⬇ ZIP
  1{
  2  "bench_version": 1,
  3  "track": "tabular",
  4  "model": "mlp_med",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.003
 10    },
 11    "sweep": [
 12      {
 13        "cfg": {
 14          "lr": 0.001
 15        },
 16        "mean": 6.546770215034485
 17      },
 18      {
 19        "cfg": {
 20          "lr": 0.003
 21        },
 22        "mean": 5.133943796157837
 23      },
 24      {
 25        "cfg": {
 26          "lr": 0.01
 27        },
 28        "mean": 5.272171854972839
 29      }
 30    ],
 31    "full": {
 32      "mean": 5.413264155387878,
 33      "std": 0.7758719453955597,
 34      "per_seed": [
 35        4.240733623504639,
 36        5.491628170013428,
 37        5.039286136627197,
 38        5.764127254486084,
 39        6.573030948638916,
 40        6.121939659118652,
 41        5.759654521942139,
 42        4.315712928771973
 43      ],
 44      "n": 8
 45    }
 46  },
 47  "idea": {
 48    "best_cfg": {
 49      "lr": 0.01
 50    },
 51    "sweep": [
 52      {
 53        "cfg": {
 54          "lr": 0.001
 55        },
 56        "mean": 6.933225512504578
 57      },
 58      {
 59        "cfg": {
 60          "lr": 0.003
 61        },
 62        "mean": 5.442892074584961
 63      },
 64      {
 65        "cfg": {
 66          "lr": 0.01
 67        },
 68        "mean": 5.218377470970154
 69      }
 70    ],
 71    "full": {
 72      "mean": 5.436576008796692,
 73      "std": 0.42245263195805716,
 74      "per_seed": [
 75        5.354506015777588,
 76        4.885490417480469,
 77        5.651634693145752,
 78        4.981878757476807,
 79        5.249680995941162,
 80        6.110184669494629,
 81        6.017618656158447,
 82        5.241613864898682
 83      ],
 84      "n": 8
 85    },
 86    "mean": 5.436576008796692,
 87    "std": 0.42245263195805716,
 88    "per_seed": [
 89      5.354506015777588,
 90      4.885490417480469,
 91      5.651634693145752,
 92      4.981878757476807,
 93      5.249680995941162,
 94      6.110184669494629,
 95      6.017618656158447,
 96      5.241613864898682
 97    ],
 98    "n": 8
 99  },
100  "comparison": {
101    "delta_mean": 0.023311853408813477,
102    "idea_wins": 4,
103    "n_pairs": 8,
104    "per_seed_diffs": [
105      1.1137723922729492,
106      -0.606137752532959,
107      0.6123485565185547,
108      -0.7822484970092773,
109      -1.323349952697754,
110      -0.011754989624023438,
111      0.2579641342163086,
112      0.925900936126709
113    ],
114    "p_value": 0.94485,
115    "mde": 0.7252095318160074,
116    "mde_rel_pct": 13.396899005828097,
117    "verdict": "no measurable effect",
118    "system_worked": false
119  },
120  "mechanism_signature": {
121    "claim": "trained ICNN predictions satisfy midpoint Jensen convexity",
122    "predicted_max_violation": 0.0,
123    "observed_max_violation": 3.814697265625e-06,
124    "observed_violation_rate": 0.0,
125    "confirmed": true,
126    "n_models": 8
127  },
128  "protocol_notes": {
129    "structural_match": "tabular is the built-in track for architecture/regularization interventions",
130    "paired_seeds": [
131      0,
132      1,
133      2,
134      3,
135      4,
136      5,
137      6,
138      7
139    ],
140    "epochs": 25,
141    "batch": 128,
142    "baseline_and_idea_share_grid": true,
143    "baseline_architecture": "2-layer width-32 ReLU MLP",
144    "idea_architecture": "2-layer width-32 ICNN with softplus W>=0 and nonnegative output weights"
145  }
146}