1{
  2  "bench_version": 1,
  3  "track": "two_view_gauge_localization_v2",
  4  "model": "mlp_tiny",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.01,
 10      "epochs": 18
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "epochs": 18
 17        },
 18        "mean": 1.3287881910800934
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "epochs": 18
 24        },
 25        "mean": 1.2643743455410004
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.01,
 30          "epochs": 18
 31        },
 32        "mean": 1.1665648221969604
 33      }
 34    ],
 35    "full": {
 36      "mean": 1.1336221471428871,
 37      "std": 0.09147791521994385,
 38      "per_seed": [
 39        1.1263655424118042,
 40        1.207707166671753,
 41        1.2102880477905273,
 42        1.1218985319137573,
 43        1.037619948387146,
 44        0.9893431067466736,
 45        1.28654146194458,
 46        1.0892133712768555
 47      ],
 48      "n": 8
 49    }
 50  },
 51  "idea": {
 52    "config": {
 53      "lr": 0.01,
 54      "epochs": 18
 55    },
 56    "mean": 1.0666468143463135,
 57    "std": 0.06930290941504633,
 58    "per_seed": [
 59      0.9820091128349304,
 60      1.1590790748596191,
 61      1.1329891681671143,
 62      1.1253324747085571,
 63      0.984858512878418,
 64      0.9815046191215515,
 65      1.096936583518982,
 66      1.0704649686813354
 67    ],
 68    "n": 8
 69  },
 70  "comparison": {
 71    "delta_mean": -0.06697533279657364,
 72    "idea_wins": 7,
 73    "n_pairs": 8,
 74    "per_seed_diffs": [
 75      -0.14435642957687378,
 76      -0.04862809181213379,
 77      -0.07729887962341309,
 78      0.0034339427947998047,
 79      -0.05276143550872803,
 80      -0.00783848762512207,
 81      -0.18960487842559814,
 82      -0.01874840259552002
 83    ],
 84    "p_value": 0.0152,
 85    "mde": 0.056931783847403286,
 86    "mde_rel_pct": 5.022112878695138,
 87    "verdict": "idea better (significant)",
 88    "system_worked": true
 89  },
 90  "mechanism_signature": {
 91    "values": {
 92      "prediction": "two-view analytic gauge initialization reduces target prediction error",
 93      "analytic_test_mse_mean": 0.011936143389903009,
 94      "trained_baseline_test_mse_mean": 1.1336221620440483,
 95      "trained_idea_test_mse_mean": 1.0666468441486359,
 96      "confirmed": true
 97    },
 98    "confirmed": true
 99  },
100  "custom_track": {
101    "name": "two_view_gauge_localization_v2",
102    "file": "two_view_registered_track_v2.py",
103    "domain": "geometry/localization"
104  }
105}