1{
  2  "bench_version": 1,
  3  "track": "tabular",
  4  "model": "two_branch_mlp",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.01
 10    },
 11    "sweep": [
 12      {
 13        "cfg": {
 14          "lr": 0.001
 15        },
 16        "mean": 0.3566400483250618
 17      },
 18      {
 19        "cfg": {
 20          "lr": 0.003
 21        },
 22        "mean": 0.27394459396600723
 23      },
 24      {
 25        "cfg": {
 26          "lr": 0.01
 27        },
 28        "mean": 0.23046058416366577
 29      }
 30    ],
 31    "full": {
 32      "mean": 0.23951114527881145,
 33      "std": 0.04028734767167621,
 34      "per_seed": [
 35        0.1893874704837799,
 36        0.2472962588071823,
 37        0.2523687779903412,
 38        0.23278982937335968,
 39        0.1838403344154358,
 40        0.25920143723487854,
 41        0.3209288716316223,
 42        0.2302761822938919
 43      ],
 44      "n": 8
 45    }
 46  },
 47  "idea": {
 48    "mean": 0.3616223409771919,
 49    "std": 0.09797382371900515,
 50    "per_seed": [
 51      0.28693538904190063,
 52      0.2831035852432251,
 53      0.28846070170402527,
 54      0.3903844356536865,
 55      0.3120247721672058,
 56      0.30676016211509705,
 57      0.5750100016593933,
 58      0.4502996802330017
 59    ],
 60    "n": 8
 61  },
 62  "comparison": {
 63    "delta_mean": 0.12211119569838047,
 64    "idea_wins": 0,
 65    "n_pairs": 8,
 66    "per_seed_diffs": [
 67      0.09754791855812073,
 68      0.035807326436042786,
 69      0.03609192371368408,
 70      0.15759460628032684,
 71      0.12818443775177002,
 72      0.047558724880218506,
 73      0.254081130027771,
 74      0.2200234979391098
 75    ],
 76    "p_value": 0.0081,
 77    "mde": 0.0701771215456511,
 78    "mde_rel_pct": 29.300148627304555,
 79    "verdict": "idea worse (significant)",
 80    "system_worked": false
 81  },
 82  "idea_sweep": [
 83    {
 84      "cfg": {
 85        "lr": 0.001
 86      },
 87      "result": {
 88        "mean": 0.3616223409771919,
 89        "std": 0.09797382371900515,
 90        "per_seed": [
 91          0.28693538904190063,
 92          0.2831035852432251,
 93          0.28846070170402527,
 94          0.3903844356536865,
 95          0.3120247721672058,
 96          0.30676016211509705,
 97          0.5750100016593933,
 98          0.4502996802330017
 99        ],
100        "n": 8
101      }
102    },
103    {
104      "cfg": {
105        "lr": 0.003
106      },
107      "result": {
108        "mean": 0.36999374255537987,
109        "std": 0.07516870461499706,
110        "per_seed": [
111          0.3048875033855438,
112          0.31957462430000305,
113          0.36907637119293213,
114          0.3171716332435608,
115          0.3382593095302582,
116          0.3245992958545685,
117          0.5264177322387695,
118          0.45996347069740295
119        ],
120        "n": 8
121      }
122    },
123    {
124      "cfg": {
125        "lr": 0.01
126      },
127      "result": {
128        "mean": 5.468297071754932,
129        "std": 6.81639231650195,
130        "per_seed": [
131          0.36661669611930847,
132          1.333353877067566,
133          0.3249545991420746,
134          2.7228331565856934,
135          13.045344352722168,
136          20.14182472229004,
137          4.995171546936035,
138          0.8162776231765747
139        ],
140        "n": 8
141      }
142    }
143  ],
144  "track_justification": "Optimizer intervention: official bench assigns optimizer ideas to tabular; both systems use the same branch-swap symmetric MLP and differ only in the update rule.",
145  "mechanism_signature": {
146    "prediction": "branch-swap parity separates even and odd gradient energy",
147    "trained_model_gradient_norms_plus": 0.0739268725739474,
148    "trained_model_gradient_norms_minus": 0.05150565393582205,
149    "ratio_minus_over_plus": 0.6967108460270031,
150    "confirmed": true
151  }
152}