1{
  2  "bench_version": 1,
  3  "track": "directional_polytope",
  4  "model": "mlp_tiny",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.03,
 10      "penalty": 0.0
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.003,
 16          "penalty": 0.0
 17        },
 18        "mean": 0.0043861757148988545
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "penalty": 1.0
 24        },
 25        "mean": 0.00439288467168808
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.003,
 30          "penalty": 10.0
 31        },
 32        "mean": 0.004514919011853635
 33      },
 34      {
 35        "cfg": {
 36          "lr": 0.01,
 37          "penalty": 0.0
 38        },
 39        "mean": 0.001629675563890487
 40      },
 41      {
 42        "cfg": {
 43          "lr": 0.01,
 44          "penalty": 1.0
 45        },
 46        "mean": 0.0016628731682430953
 47      },
 48      {
 49        "cfg": {
 50          "lr": 0.01,
 51          "penalty": 10.0
 52        },
 53        "mean": 0.001958355976967141
 54      },
 55      {
 56        "cfg": {
 57          "lr": 0.03,
 58          "penalty": 0.0
 59        },
 60        "mean": 0.0009300831443397328
 61      },
 62      {
 63        "cfg": {
 64          "lr": 0.03,
 65          "penalty": 1.0
 66        },
 67        "mean": 0.0009807027818169445
 68      },
 69      {
 70        "cfg": {
 71          "lr": 0.03,
 72          "penalty": 10.0
 73        },
 74        "mean": 0.004855888721067458
 75      }
 76    ],
 77    "full": {
 78      "mean": 0.0010100424842676148,
 79      "std": 0.0001477102326352443,
 80      "per_seed": [
 81        0.0009797323727980256,
 82        0.0010062798392027617,
 83        0.0009645665995776653,
 84        0.0007697537657804787,
 85        0.001155781908892095,
 86        0.0010888965334743261,
 87        0.0008554105297662318,
 88        0.001259918324649334
 89      ],
 90      "n": 8
 91    }
 92  },
 93  "idea": {
 94    "mean": 0.001285703867324628,
 95    "std": 9.055997386539207e-05,
 96    "per_seed": [
 97      0.0013685496523976326,
 98      0.001279718242585659,
 99      0.0012515878770500422,
100      0.001286455080844462,
101      0.0013379424344748259,
102      0.001405661110766232,
103      0.0010852159466594458,
104      0.0012705005938187242
105    ],
106    "n": 8
107  },
108  "comparison": {
109    "delta_mean": 0.00027566138305701315,
110    "idea_wins": 0,
111    "n_pairs": 8,
112    "per_seed_diffs": [
113      0.000388817279599607,
114      0.0002734384033828974,
115      0.0002870212774723768,
116      0.0005167013150639832,
117      0.00018216052558273077,
118      0.0003167645772919059,
119      0.000229805416893214,
120      1.0582269169390202e-05
121    ],
122    "p_value": 0.0081,
123    "mde": 0.00012395723646933067,
124    "mde_rel_pct": 12.272477484866636,
125    "verdict": "idea worse (significant)",
126    "system_worked": false
127  },
128  "mechanism_signature": {
129    "quantity": "diagonal Jacobian of trained feasibility head",
130    "predicted_mean_abs": 2.8536648750305176,
131    "observed_mean_abs": 2.853664852678776,
132    "max_relative_error": 1.4644062872233774e-07,
133    "confirmed": true
134  },
135  "idea_sweep": [
136    {
137      "cfg": {
138        "lr": 0.03,
139        "eps": 0.05
140      },
141      "mean": 0.007740114291664213
142    },
143    {
144      "cfg": {
145        "lr": 0.03,
146        "eps": 0.1
147      },
148      "mean": 0.0024592391564510763
149    },
150    {
151      "cfg": {
152        "lr": 0.03,
153        "eps": 0.2
154      },
155      "mean": 0.001296577713219449
156    }
157  ],
158  "idea_best_cfg": {
159    "lr": 0.03,
160    "eps": 0.2
161  },
162  "runtime_seconds": 35.400410652160645,
163  "protocol_note": "Registered directional_polytope; same MLP, end-to-end trained systems; box log barrier feasibility."
164}