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": 0.12,
 10      "epochs": 15
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.03,
 16          "epochs": 15
 17        },
 18        "mean": 17.87290906906128
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.06,
 23          "epochs": 15
 24        },
 25        "mean": 18.236158847808838
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.12,
 30          "epochs": 15
 31        },
 32        "mean": 7.929520010948181
 33      }
 34    ],
 35    "full": {
 36      "mean": 12.419063031673431,
 37      "std": 4.653164949772833,
 38      "per_seed": [
 39        7.232073783874512,
 40        9.054744720458984,
 41        8.681772232055664,
 42        6.7494893074035645,
 43        14.7591552734375,
 44        18.111841201782227,
 45        18.314743041992188,
 46        16.448684692382812
 47      ],
 48      "n": 8
 49    }
 50  },
 51  "idea": {
 52    "mean": 12.419063031673431,
 53    "std": 4.653164949772833,
 54    "per_seed": [
 55      7.232073783874512,
 56      9.054744720458984,
 57      8.681772232055664,
 58      6.7494893074035645,
 59      14.7591552734375,
 60      18.111841201782227,
 61      18.314743041992188,
 62      16.448684692382812
 63    ],
 64    "n": 8,
 65    "best_cfg": {
 66      "lr": 0.12,
 67      "epochs": 15
 68    },
 69    "sweep": [
 70      {
 71        "cfg": {
 72          "lr": 0.03,
 73          "epochs": 15
 74        },
 75        "mean": 22.84104561805725,
 76        "std": 11.555765903919722,
 77        "per_seed": [
 78          18.62317657470703,
 79          17.563692092895508,
 80          16.379343032836914,
 81          13.706111907958984,
 82          13.296585083007812,
 83          38.26602554321289,
 84          46.30095291137695,
 85          18.592477798461914
 86        ],
 87        "n": 8
 88      },
 89      {
 90        "cfg": {
 91          "lr": 0.06,
 92          "epochs": 15
 93        },
 94        "mean": 17.78099489212036,
 95        "std": 4.228547463617409,
 96        "per_seed": [
 97          9.133557319641113,
 98          17.798871994018555,
 99          16.796100616455078,
100          15.525813102722168,
101          24.372446060180664,
102          17.10875129699707,
103          21.30739402770996,
104          20.20502471923828
105        ],
106        "n": 8
107      },
108      {
109        "cfg": {
110          "lr": 0.12,
111          "epochs": 15
112        },
113        "mean": 12.419063031673431,
114        "std": 4.653164949772833,
115        "per_seed": [
116          7.232073783874512,
117          9.054744720458984,
118          8.681772232055664,
119          6.7494893074035645,
120          14.7591552734375,
121          18.111841201782227,
122          18.314743041992188,
123          16.448684692382812
124        ],
125        "n": 8,
126        "best_cfg": {
127          "lr": 0.12,
128          "epochs": 15
129        }
130      }
131    ]
132  },
133  "comparison": {
134    "delta_mean": 0.0,
135    "idea_wins": 0,
136    "n_pairs": 8,
137    "per_seed_diffs": [
138      0.0,
139      0.0,
140      0.0,
141      0.0,
142      0.0,
143      0.0,
144      0.0,
145      0.0
146    ],
147    "p_value": 1.0,
148    "mde": 0.0,
149    "mde_rel_pct": 0.0,
150    "verdict": "no measurable effect",
151    "system_worked": false
152  },
153  "mechanism_signature": {
154    "prediction": "negative secant pairs are corrected to positive curvature; correction should occur only when the trained consecutive-gradient secant is negative",
155    "baseline_observed": {
156      "samples": 128,
157      "mean_abs_parameter_input_gradient": 2.754227876663208,
158      "finite_fraction": 1.0,
159      "output_std": 3.702409267425537
160    },
161    "idea_observed": {
162      "samples": 128,
163      "mean_abs_parameter_input_gradient": 2.754227876663208,
164      "finite_fraction": 1.0,
165      "output_std": 3.702409267425537
166    },
167    "trained_pair_counts": {
168      "baseline_negative_pairs": 0,
169      "idea_negative_pairs": 0,
170      "idea_corrected_pairs": 0
171    },
172    "confirmed": false,
173    "measurement_note": "Behavior probes and pair counts are measured from trained official-tabular systems; no analytical toy identity is used."
174  },
175  "protocol_note": "Official registered tabular track; optimizer-only custom loop because the intervention changes training."
176}