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": 0.01,
 10      "scale": 1.3
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "scale": 0.7
 17        },
 18        "mean": 17.77549433708191
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "scale": 1.0
 24        },
 25        "mean": 13.713647603988647
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.01,
 30          "scale": 1.3
 31        },
 32        "mean": 7.7490280866622925
 33      }
 34    ],
 35    "full": {
 36      "mean": 7.927547872066498,
 37      "std": 0.6638450060890482,
 38      "per_seed": [
 39        6.863780498504639,
 40        8.03250503540039,
 41        8.228401184082031,
 42        7.871425628662109,
 43        6.849928855895996,
 44        8.388378143310547,
 45        8.470486640930176,
 46        8.715476989746094
 47      ],
 48      "n": 8
 49    }
 50  },
 51  "idea": {
 52    "cfg": {
 53      "lr": 0.01,
 54      "scale": 1.3
 55    },
 56    "mean": 9.446731090545654,
 57    "std": 0.9494091152353562,
 58    "per_seed": [
 59      10.052746772766113,
 60      10.37270736694336,
 61      8.832763671875,
 62      8.52699089050293,
 63      7.573822021484375,
 64      10.16633129119873,
 65      10.212395668029785,
 66      9.836091041564941
 67    ],
 68    "n": 8
 69  },
 70  "comparison": {
 71    "delta_mean": 1.5191832184791565,
 72    "idea_wins": 0,
 73    "n_pairs": 8,
 74    "per_seed_diffs": [
 75      3.1889662742614746,
 76      2.3402023315429688,
 77      0.6043624877929688,
 78      0.6555652618408203,
 79      0.7238931655883789,
 80      1.7779531478881836,
 81      1.7419090270996094,
 82      1.1206140518188477
 83    ],
 84    "p_value": 0.0081,
 85    "mde": 0.7710830795135443,
 86    "mde_rel_pct": 9.726627854629957,
 87    "verdict": "idea worse (significant)",
 88    "system_worked": false
 89  },
 90  "track_choice": "tabular: initialization is explicitly covered by the benchmark mapping; same mlp_tiny and training loop used.",
 91  "idea_sweep": [
 92    {
 93      "cfg": {
 94        "lr": 0.01,
 95        "scale": 1.3
 96      }
 97    },
 98    {
 99      "cfg": {
100        "lr": 0.001,
101        "scale": 0.7
102      }
103    }
104  ],
105  "trained_geometry": {
106    "baseline_median_nn": 0.4159104786813259,
107    "idea_median_nn": 0.06890426948666573,
108    "baseline_mean_nn": 0.4142976365983486,
109    "idea_mean_nn": 0.07547109667211771
110  },
111  "mechanism_signature": {
112    "quantity": "trained first-layer median nearest-neighbor distance",
113    "prediction": "Coulomb repulsion preserves larger inter-row spacing as N=64 bank grows",
114    "baseline_observed": 0.4159104786813259,
115    "idea_observed": 0.06890426948666573,
116    "observed_ratio_idea_over_baseline": 0.1656709148207366,
117    "confirmed": false,
118    "n_models": 8
119  }
120}