1{
  2  "bench_version": 1,
  3  "track": "masked_blur_inverse",
  4  "model": "mlp_tiny",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.003,
 10      "lambda": 0.0
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "lambda": 0.0
 17        },
 18        "mean": 0.027918022125959396
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "lambda": 0.0
 24        },
 25        "mean": 0.021377066150307655
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.01,
 30          "lambda": 0.0
 31        },
 32        "mean": 0.021512716310098767
 33      }
 34    ],
 35    "full": {
 36      "mean": 0.021377066150307655,
 37      "std": 0.0003192543389189106,
 38      "per_seed": [
 39        0.02174585871398449,
 40        0.021633073687553406,
 41        0.02161465398967266,
 42        0.021050795912742615,
 43        0.020985819399356842,
 44        0.020888708531856537,
 45        0.021562591195106506,
 46        0.021535027772188187
 47      ],
 48      "n": 8
 49    }
 50  },
 51  "idea": {
 52    "mean": 0.02130480553023517,
 53    "std": 0.0003312501842748453,
 54    "per_seed": [
 55      0.021708831191062927,
 56      0.02170940674841404,
 57      0.021566448733210564,
 58      0.021123606711626053,
 59      0.020895397290587425,
 60      0.020775478333234787,
 61      0.0213462021201849,
 62      0.021313073113560677
 63    ],
 64    "n": 8
 65  },
 66  "comparison": {
 67    "delta_mean": -7.226062007248402e-05,
 68    "idea_wins": 6,
 69    "n_pairs": 8,
 70    "per_seed_diffs": [
 71      -3.7027522921562195e-05,
 72      7.633306086063385e-05,
 73      -4.820525646209717e-05,
 74      7.281079888343811e-05,
 75      -9.042210876941681e-05,
 76      -0.00011323019862174988,
 77      -0.00021638907492160797,
 78      -0.00022195465862751007
 79    ],
 80    "p_value": 0.1159,
 81    "mde": 9.48822665875107e-05,
 82    "mde_rel_pct": 0.4438507413522934,
 83    "verdict": "no measurable effect",
 84    "system_worked": false
 85  },
 86  "custom_track": {
 87    "name": "masked_blur_inverse",
 88    "file": "custom_inverse_track.py",
 89    "domain": "pde_inverse"
 90  },
 91  "mathcheck": {
 92    "max_relative_column_fd_error": 4.637162161899401e-16,
 93    "weight_min": 0.609901252345,
 94    "weight_max": 1.5739333755883023,
 95    "weight_mean": 1.0
 96  },
 97  "idea_sweep": [
 98    {
 99      "cfg": {
100        "lr": 0.001,
101        "lambda": 0.002
102      },
103      "mean": 0.027870018035173416
104    },
105    {
106      "cfg": {
107        "lr": 0.003,
108        "lambda": 0.002
109      },
110      "mean": 0.02130480553023517
111    },
112    {
113      "cfg": {
114        "lr": 0.01,
115        "lambda": 0.002
116      },
117      "mean": 0.021524549229070544
118    }
119  ],
120  "mechanism_signature": {
121    "predicted": "weighted TV should alter spatial gradient allocation in proportion to cached sensitivity",
122    "observed_mean_horizontal_gradient": 0.09227584861218929,
123    "baseline_mean_horizontal_gradient": 0.09304057341068983,
124    "relative_change": -0.00821926145193641,
125    "confirmed": true
126  }
127}