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": 12
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.0015,
 16          "epochs": 12
 17        },
 18        "mean": 51.05318737030029
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "epochs": 12
 24        },
 25        "mean": 17.679327249526978
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.006,
 30          "epochs": 12
 31        },
 32        "mean": 15.346336364746094
 33      }
 34    ],
 35    "full": {
 36      "mean": 15.31271517276764,
 37      "std": 0.822924635150327,
 38      "per_seed": [
 39        15.071464538574219,
 40        14.645296096801758,
 41        16.341999053955078,
 42        14.45390510559082,
 43        14.129950523376465,
 44        15.382646560668945,
 45        16.234203338623047,
 46        16.24225616455078
 47      ],
 48      "n": 8
 49    }
 50  },
 51  "idea": {
 52    "mean": 103.05307102203369,
 53    "std": 20.493133534705294,
 54    "per_seed": [
 55      130.216796875,
 56      99.72685241699219,
 57      92.01232147216797,
 58      98.55255889892578,
 59      89.51197052001953,
 60      66.24312591552734,
 61      119.2832260131836,
 62      128.87771606445312
 63    ],
 64    "n": 8,
 65    "sweep": [
 66      {
 67        "cfg": {
 68          "lr": 0.0015,
 69          "epochs": 12
 70        },
 71        "mean": 218.74835777282715,
 72        "std": 10.794174158154743,
 73        "per_seed": [
 74          230.59194946289062,
 75          216.66702270507812,
 76          231.0927276611328,
 77          194.62643432617188,
 78          222.1121063232422,
 79          216.87411499023438,
 80          214.635498046875,
 81          223.3870086669922
 82        ],
 83        "n": 8
 84      },
 85      {
 86        "cfg": {
 87          "lr": 0.003,
 88          "epochs": 12
 89        },
 90        "mean": 197.7052059173584,
 91        "std": 8.29119691504502,
 92        "per_seed": [
 93          193.0837860107422,
 94          201.17518615722656,
 95          196.46839904785156,
 96          183.23458862304688,
 97          213.4451141357422,
 98          198.17083740234375,
 99          203.3914337158203,
100          192.67230224609375
101        ],
102        "n": 8
103      },
104      {
105        "cfg": {
106          "lr": 0.006,
107          "epochs": 12
108        },
109        "mean": 103.05307102203369,
110        "std": 20.493133534705294,
111        "per_seed": [
112          130.216796875,
113          99.72685241699219,
114          92.01232147216797,
115          98.55255889892578,
116          89.51197052001953,
117          66.24312591552734,
118          119.2832260131836,
119          128.87771606445312
120        ],
121        "n": 8
122      }
123    ]
124  },
125  "comparison": {
126    "delta_mean": 87.74035584926605,
127    "idea_wins": 0,
128    "n_pairs": 8,
129    "per_seed_diffs": [
130      115.14533233642578,
131      85.08155632019043,
132      75.67032241821289,
133      84.09865379333496,
134      75.38201999664307,
135      50.8604793548584,
136      103.04902267456055,
137      112.63545989990234
138    ],
139    "p_value": 0.0081,
140    "mde": 18.110913310349133,
141    "mde_rel_pct": 118.27369023723402,
142    "verdict": "idea worse (significant)",
143    "system_worked": false,
144    "paired_delta_mean": 87.74035584926605,
145    "permutation_pvalue": 0.0081
146  },
147  "mechanism_signature": {
148    "prediction": "greedy selected scenarios cover the high-loss tail",
149    "observed_selected_tail_fraction": 0.9,
150    "observed_baseline_top_tail_fraction": 1.0,
151    "selected_count": 20,
152    "idea_train_mse": 100.62527465820312,
153    "baseline_train_mse": 15.9457426071167,
154    "confirmed": true
155  },
156  "core_math_check": {
157    "proxy_epsilon_N_100_200_400_800": [
158      0.1351292546497023,
159      0.06756462732485115,
160      0.033782313662425574,
161      0.016891156831212787
162    ],
163    "monotone_decrease": true,
164    "N": 400,
165    "k": 20,
166    "beta": 1e-05
167  },
168  "bench_report": {
169    "track_justification": "tabular is the harness-matched track for optimizer/training-dynamics/calibration interventions",
170    "custom_track": null,
171    "protocol": "8 paired seeds; baseline sweep and idea sweep share the exact lr/epoch union",
172    "budget_note": "small 400-sample, 12-epoch MLP; compression uses active-set retraining and final retraining"
173  }
174}