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    },
 11    "sweep": [
 12      {
 13        "cfg": {
 14          "lr": 0.001
 15        },
 16        "mean": 17.482568740844727
 17      },
 18      {
 19        "cfg": {
 20          "lr": 0.003
 21        },
 22        "mean": 12.535736083984375
 23      },
 24      {
 25        "cfg": {
 26          "lr": 0.006
 27        },
 28        "mean": 9.091599702835083
 29      }
 30    ],
 31    "full": {
 32      "mean": 9.481420636177063,
 33      "std": 0.9072998646400672,
 34      "per_seed": [
 35        8.627572059631348,
 36        8.736048698425293,
 37        8.82056999206543,
 38        10.182208061218262,
 39        8.84079647064209,
 40        9.23827075958252,
 41        11.380922317504883,
 42        10.02497673034668
 43      ],
 44      "n": 8
 45    }
 46  },
 47  "idea": {
 48    "mean": 9.610056400299072,
 49    "std": 1.0356528912793441,
 50    "per_seed": [
 51      9.995634078979492,
 52      8.813212394714355,
 53      8.717215538024902,
 54      9.493178367614746,
 55      8.498827934265137,
 56      9.12738037109375,
 57      11.815202713012695,
 58      10.4197998046875
 59    ],
 60    "n": 8
 61  },
 62  "comparison": {
 63    "delta_mean": 0.12863576412200928,
 64    "idea_wins": 4,
 65    "n_pairs": 8,
 66    "per_seed_diffs": [
 67      1.3680620193481445,
 68      0.0771636962890625,
 69      -0.10335445404052734,
 70      -0.6890296936035156,
 71      -0.3419685363769531,
 72      -0.11089038848876953,
 73      0.4342803955078125,
 74      0.3948230743408203
 75    ],
 76    "p_value": 0.61945,
 77    "mde": 0.5197178870759055,
 78    "mde_rel_pct": 5.481434766145522,
 79    "verdict": "no significant win",
 80    "system_worked": false
 81  },
 82  "mechanism_signature": {
 83    "signature": {
 84      "quantity": "family allocation variance proxy sum(u_c^2/n_c)",
 85      "predicted_mean": 0.0982159124687314,
 86      "observed_mean": 0.09136932715773582,
 87      "relative_error": 0.06970953218171544,
 88      "confirmed": true,
 89      "note": "u and n are computed from trained model residuals; observed proxy uses held-out family residual variances."
 90    },
 91    "idea_sweep": [
 92      {
 93        "cfg": {
 94          "lr": 0.001,
 95          "epsilon": 0.05
 96        },
 97        "result": {
 98          "mean": 18.78651201725006,
 99          "std": 2.1210090381527644,
100          "per_seed": [
101            17.013912200927734,
102            15.361128807067871,
103            21.161161422729492,
104            16.918540954589844,
105            19.89978790283203,
106            17.97165870666504,
107            20.517269134521484,
108            21.448637008666992
109          ],
110          "n": 8
111        }
112      },
113      {
114        "cfg": {
115          "lr": 0.003,
116          "epsilon": 0.15
117        },
118        "result": {
119          "mean": 13.162447214126587,
120          "std": 1.3889931613668718,
121          "per_seed": [
122            13.541401863098145,
123            11.67493724822998,
124            12.949481964111328,
125            12.458735466003418,
126            11.737956047058105,
127            12.227039337158203,
128            15.25829792022705,
129            15.451727867126465
130          ],
131          "n": 8
132        }
133      },
134      {
135        "cfg": {
136          "lr": 0.006,
137          "epsilon": 0.3
138        },
139        "result": {
140          "mean": 9.610056400299072,
141          "std": 1.0356528912793441,
142          "per_seed": [
143            9.995634078979492,
144            8.813212394714355,
145            8.717215538024902,
146            9.493178367614746,
147            8.498827934265137,
148            9.12738037109375,
149            11.815202713012695,
150            10.4197998046875
151          ],
152          "n": 8
153        }
154      }
155    ],
156    "track_rationale": "Family-conditioned minibatch selection is an optimizer/training-data mechanism; tabular regression is the mandated structural track for optimizer/data mechanics."
157  },
158  "stage2_protocol": {
159    "paired_seeds": [
160      0,
161      1,
162      2,
163      3,
164      4,
165      5,
166      6,
167      7
168    ],
169    "epochs": 12,
170    "batch": 64,
171    "baseline_grid": [
172      {
173        "lr": 0.001
174      },
175      {
176        "lr": 0.003
177      },
178      {
179        "lr": 0.006
180      }
181    ],
182    "idea_grid": [
183      {
184        "lr": 0.001,
185        "epsilon": 0.05
186      },
187      {
188        "lr": 0.003,
189        "epsilon": 0.15
190      },
191      {
192        "lr": 0.006,
193        "epsilon": 0.3
194      }
195    ]
196  }
197}