1{
  2  "bench_version": 1,
  3  "track": "piola_surface_operator",
  4  "model": "mlp_tiny",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.003
 10    },
 11    "sweep": [
 12      {
 13        "cfg": {
 14          "lr": 0.001
 15        },
 16        "mean": 0.398487139493227
 17      },
 18      {
 19        "cfg": {
 20          "lr": 0.003
 21        },
 22        "mean": 0.3586462587118149
 23      },
 24      {
 25        "cfg": {
 26          "lr": 0.01
 27        },
 28        "mean": 0.46151376888155937
 29      }
 30    ],
 31    "full": {
 32      "mean": 0.33389632031321526,
 33      "std": 0.010530313211265758,
 34      "per_seed": [
 35        0.32944396138191223,
 36        0.3304601013660431,
 37        0.3288800120353699,
 38        0.32970303297042847,
 39        0.3304898738861084,
 40        0.3331677317619324,
 41        0.36147114634513855,
 42        0.32755470275878906
 43      ],
 44      "n": 8
 45    },
 46    "parity_grid": [
 47      {
 48        "lr": 0.001
 49      },
 50      {
 51        "lr": 0.003
 52      },
 53      {
 54        "lr": 0.01
 55      }
 56    ],
 57    "selected_lr": 0.003,
 58    "idea_grid": [
 59      {
 60        "cfg": {
 61          "lr": 0.001
 62        },
 63        "mean": 0.33370736241340637
 64      },
 65      {
 66        "cfg": {
 67          "lr": 0.003
 68        },
 69        "mean": 0.24209699407219887
 70      },
 71      {
 72        "cfg": {
 73          "lr": 0.01
 74        },
 75        "mean": 0.2798946388065815
 76      }
 77    ]
 78  },
 79  "idea": {
 80    "mean": 0.24209699407219887,
 81    "std": 0.007792167833569856,
 82    "per_seed": [
 83      0.23542946577072144,
 84      0.23528780043125153,
 85      0.23226389288902283,
 86      0.2527761161327362,
 87      0.23969487845897675,
 88      0.24286238849163055,
 89      0.2553972601890564,
 90      0.24306415021419525
 91    ],
 92    "n": 8
 93  },
 94  "comparison": {
 95    "delta_mean": -0.09179932624101639,
 96    "idea_wins": 8,
 97    "n_pairs": 8,
 98    "per_seed_diffs": [
 99      -0.0940144956111908,
100      -0.09517230093479156,
101      -0.09661611914634705,
102      -0.07692691683769226,
103      -0.09079499542713165,
104      -0.09030534327030182,
105      -0.10607388615608215,
106      -0.08449055254459381
107    ],
108    "p_value": 0.0081,
109    "mde": 0.007214329180313888,
110    "mde_rel_pct": 2.160649501481898,
111    "verdict": "idea better (significant)",
112    "system_worked": true
113  },
114  "custom_track": {
115    "name": "piola_surface_operator",
116    "file": "piola_surface_operator.py",
117    "domain": "pde"
118  },
119  "protocol_notes": {
120    "epochs": 18,
121    "batch": 64,
122    "seeds": [
123      11,
124      29,
125      47,
126      71,
127      89,
128      107,
129      131,
130      149
131    ],
132    "baseline_and_idea_same_architecture": true,
133    "baseline_and_idea_same_lr_union": true
134  },
135  "mechanism_signature": {
136    "quantity": "trained reference prediction RMSE and cross-geometry output consistency",
137    "baseline_observed_reference_prediction_rmse": 0.5627477765083313,
138    "idea_observed_reference_prediction_rmse": 0.4821018874645233,
139    "baseline_geometry_output_consistency": 0.31939321465817877,
140    "idea_geometry_output_consistency": 0.03450808376451315,
141    "predicted_effect": "inverse-Piola inputs should reduce geometry-induced variation in reference predictions",
142    "confirmed": true
143  }
144}