1{
  2  "baseline": {
  3    "best_cfg": {
  4      "epochs": 35,
  5      "lr": 0.01,
  6      "width": 48
  7    },
  8    "full": {
  9      "mean": 0.1913430094718933,
 10      "n": 8,
 11      "per_seed": [
 12        0.21796821057796478,
 13        0.18948782980442047,
 14        0.18038326501846313,
 15        0.17539961636066437,
 16        0.2016671746969223,
 17        0.19406883418560028,
 18        0.18906868994235992,
 19        0.18270045518875122
 20      ],
 21      "std": 0.01265175674434216
 22    },
 23    "sweep": [
 24      {
 25        "cfg": {
 26          "epochs": 35,
 27          "lr": 0.001,
 28          "width": 48
 29        },
 30        "mean": 0.2922794222831726
 31      },
 32      {
 33        "cfg": {
 34          "epochs": 35,
 35          "lr": 0.003,
 36          "width": 48
 37        },
 38        "mean": 0.2332044392824173
 39      },
 40      {
 41        "cfg": {
 42          "epochs": 35,
 43          "lr": 0.01,
 44          "width": 48
 45        },
 46        "mean": 0.1908097304403782
 47      }
 48    ]
 49  },
 50  "bench_version": 1,
 51  "comparison": {
 52    "delta_mean": -0.10181683763221372,
 53    "idea_wins": 6,
 54    "mde": 0.08970938035328954,
 55    "mde_rel_pct": 46.88406469663429,
 56    "n_pairs": 8,
 57    "p_value": 0.03805,
 58    "per_seed_diffs": [
 59      -0.13966041058301926,
 60      -0.18819955282378942,
 61      -0.1548399105668068,
 62      -0.17267675884068012,
 63      0.06887666881084442,
 64      0.06751273572444916,
 65      -0.11764338612556458,
 66      -0.17790408665314317
 67    ],
 68    "system_worked": true,
 69    "verdict": "idea better (significant)"
 70  },
 71  "custom_track": {
 72    "domain": "graph",
 73    "file": "holonomy_track.py",
 74    "name": "directed_path_holonomy"
 75  },
 76  "idea": {
 77    "mean": 0.08952617183967959,
 78    "n": 8,
 79    "per_seed": [
 80      0.07830779999494553,
 81      0.0012882769806310534,
 82      0.02554335445165634,
 83      0.0027228575199842453,
 84      0.2705438435077667,
 85      0.26158156991004944,
 86      0.07142530381679535,
 87      0.004796368535608053
 88    ],
 89    "std": 0.10572229702516253
 90  },
 91  "idea_best_cfg": {
 92    "epochs": 35,
 93    "lr": 0.01,
 94    "width": 48
 95  },
 96  "idea_sweep": [
 97    {
 98      "cfg": {
 99        "epochs": 35,
100        "lr": 0.001,
101        "width": 48
102      },
103      "mean": 0.33307167142629623
104    },
105    {
106      "cfg": {
107        "epochs": 35,
108        "lr": 0.003,
109        "width": 48
110      },
111      "mean": 0.1860583908855915
112    },
113    {
114      "cfg": {
115        "epochs": 35,
116        "lr": 0.01,
117        "width": 48
118      },
119      "mean": 0.02696557223680429
120    }
121  ],
122  "mechanism_signature": {
123    "confirmed": false,
124    "observed_mean": {
125      "baseline_test_rmse": 0.4371945932507515,
126      "idea_test_rmse": 0.2369789406657219,
127      "order_sensitivity": 0.36837229039520025,
128      "system_gap": 0.3969488888978958
129    },
130    "prediction": "ordered holonomy should retain order: reversal sensitivity exceeds additive baseline gap"
131  },
132  "metric_direction": "lower is better",
133  "model": "matched_matrix_mlp",
134  "n_seeds": 8,
135  "notes": "Baseline and idea share edge encoder and head; only sum versus ordered matrix product differs.",
136  "track": "directed_path_holonomy"
137}