1{
  2  "bench_version": 1,
  3  "track": "spd_covariance_classification",
  4  "model": "mlp_tiny",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.001,
 10      "epochs": 18
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "epochs": 18
 17        },
 18        "mean": 0.004999999888241291
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "epochs": 18
 24        },
 25        "mean": 0.006249999860301614
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.01,
 30          "epochs": 18
 31        },
 32        "mean": 0.007499999832361937
 33      }
 34    ],
 35    "full": {
 36      "mean": 0.005624999874271452,
 37      "std": 0.0029973946350734574,
 38      "per_seed": [
 39        0.004999999888241291,
 40        0.009999999776482582,
 41        0.004999999888241291,
 42        0.0,
 43        0.009999999776482582,
 44        0.004999999888241291,
 45        0.004999999888241291,
 46        0.004999999888241291
 47      ],
 48      "n": 8
 49    }
 50  },
 51  "idea": {
 52    "mean": 0.25124998565297574,
 53    "std": 0.32049520383799557,
 54    "per_seed": [
 55      0.004999999888241291,
 56      0.004999999888241291,
 57      0.6649999618530273,
 58      0.004999999888241291,
 59      0.0,
 60      0.6649999618530273,
 61      0.6649999618530273,
 62      0.0
 63    ],
 64    "n": 8
 65  },
 66  "comparison": {
 67    "delta_mean": 0.24562498577870429,
 68    "idea_wins": 3,
 69    "n_pairs": 8,
 70    "per_seed_diffs": [
 71      0.0,
 72      -0.004999999888241291,
 73      0.659999961964786,
 74      0.004999999888241291,
 75      -0.009999999776482582,
 76      0.659999961964786,
 77      0.659999961964786,
 78      -0.004999999888241291
 79    ],
 80    "p_value": 0.2308,
 81    "mde": 0.28693631664685193,
 82    "mde_rel_pct": 5101.090187740063,
 83    "verdict": "no significant win",
 84    "system_worked": false
 85  },
 86  "mechanism_signature": {
 87    "prediction": "factor-generated prototypes remain SPD and logits finite after NN training",
 88    "predicted_min_eigenvalue_bound": 0.0,
 89    "observed_min_prototype_eigenvalue": 0.42081671953201294,
 90    "observed_finite_logit_fraction": 1.0,
 91    "confirmed": true,
 92    "idea_settings_tried": [
 93      {
 94        "lr": 0.001,
 95        "epochs": 18
 96      },
 97      {
 98        "lr": 0.003,
 99        "epochs": 18
100      },
101      {
102        "lr": 0.01,
103        "epochs": 18
104      }
105    ]
106  },
107  "idea_selected_cfg": {
108    "lr": 0.003,
109    "epochs": 18
110  }
111}