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      "epochs": 30
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "epochs": 30
 17        },
 18        "mean": 6.278503656387329
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "epochs": 30
 24        },
 25        "mean": 3.3817883133888245
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.006,
 30          "epochs": 30
 31        },
 32        "mean": 3.0755263566970825
 33      }
 34    ],
 35    "full": {
 36      "mean": 3.486333578824997,
 37      "std": 0.633233116371069,
 38      "per_seed": [
 39        2.968747138977051,
 40        2.9661331176757812,
 41        3.391489028930664,
 42        2.975736141204834,
 43        4.913453578948975,
 44        3.106180191040039,
 45        3.927138328552246,
 46        3.6417911052703857
 47      ],
 48      "n": 8
 49    }
 50  },
 51  "idea": {
 52    "mean": 1.7247163504362106,
 53    "std": 0.2112279864426216,
 54    "per_seed": [
 55      1.729775071144104,
 56      1.700377106666565,
 57      1.6194583177566528,
 58      1.3495277166366577,
 59      1.6963016986846924,
 60      1.6350337266921997,
 61      2.0824673175811768,
 62      1.9847898483276367
 63    ],
 64    "n": 8
 65  },
 66  "comparison": {
 67    "delta_mean": -1.7616172283887863,
 68    "idea_wins": 8,
 69    "n_pairs": 8,
 70    "per_seed_diffs": [
 71      -1.2389720678329468,
 72      -1.2657560110092163,
 73      -1.7720307111740112,
 74      -1.6262084245681763,
 75      -3.2171518802642822,
 76      -1.4711464643478394,
 77      -1.8446710109710693,
 78      -1.657001256942749
 79    ],
 80    "p_value": 0.0081,
 81    "mde": 0.5248752320909208,
 82    "mde_rel_pct": 15.055221200830132,
 83    "verdict": "idea better (significant)",
 84    "system_worked": true
 85  },
 86  "idea_sweep": [
 87    {
 88      "cfg": {
 89        "lr": 0.001,
 90        "epochs": 30
 91      },
 92      "mean": 3.4459262788295746,
 93      "std": 0.2142069786440496,
 94      "per_seed": [
 95        3.629073143005371,
 96        3.5230846405029297,
 97        2.9594202041625977,
 98        3.4428420066833496,
 99        3.6288678646087646,
100        3.5347254276275635,
101        3.2687673568725586,
102        3.580629587173462
103      ],
104      "n": 8
105    },
106    {
107      "cfg": {
108        "lr": 0.003,
109        "epochs": 30
110      },
111      "mean": 1.7938710153102875,
112      "std": 0.18505377667547435,
113      "per_seed": [
114        1.642995834350586,
115        2.02911639213562,
116        1.8970131874084473,
117        1.7373865842819214,
118        1.4786462783813477,
119        1.8234977722167969,
120        2.0580201148986816,
121        1.684291958808899
122      ],
123      "n": 8
124    },
125    {
126      "cfg": {
127        "lr": 0.006,
128        "epochs": 30
129      },
130      "mean": 1.7247163504362106,
131      "std": 0.2112279864426216,
132      "per_seed": [
133        1.729775071144104,
134        1.700377106666565,
135        1.6194583177566528,
136        1.3495277166366577,
137        1.6963016986846924,
138        1.6350337266921997,
139        2.0824673175811768,
140        1.9847898483276367
141      ],
142      "n": 8
143    }
144  ],
145  "matched_structure": "Friedman#1 tabular regression; regularization/MLP intervention",
146  "hyperparameter_parity": "baseline and idea both evaluated lr={0.001,0.003,0.006}, epochs=30",
147  "selected_idea_cfg": {
148    "lr": 0.006,
149    "epochs": 30
150  },
151  "mechanism_signature": {
152    "trained_test_mse": 1.729775071144104,
153    "free_branch_mse": 16.04800796508789,
154    "geom_branch_mse": 71.31401824951172,
155    "mean_alpha": 0.2708725333213806,
156    "alpha_std": 0.17698463797569275,
157    "alpha_support_corr": -0.01807535826279521,
158    "initial_alpha_at_beta_minus4": 0.01798621006309986,
159    "predicted": {
160      "learned_gate_nonconstant": true,
161      "prior_use_should_be_suppressible": true
162    },
163    "confirmed": true
164  }
165}