1{
  2  "bench_version": 1,
  3  "track": "poisson_dirichlet",
  4  "model": "mlp_tiny",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.01,
 10      "weight_decay": 0.0
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "weight_decay": 0.0
 17        },
 18        "mean": 0.0014014299667906016
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "weight_decay": 0.0
 24        },
 25        "mean": 0.00047322634054580703
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.01,
 30          "weight_decay": 0.0
 31        },
 32        "mean": 0.00023210687140817754
 33      }
 34    ],
 35    "full": {
 36      "mean": 0.00022453273413702846,
 37      "std": 2.826627803497815e-05,
 38      "per_seed": [
 39        0.00021784205455332994,
 40        0.00023001956287771463,
 41        0.00027551333187147975,
 42        0.00020505253633018583,
 43        0.00025107755209319293,
 44        0.0002373710594838485,
 45        0.00019424316997174174,
 46        0.0001851426059147343
 47      ],
 48      "n": 8
 49    },
 50    "idea_union_trials": [
 51      {
 52        "cfg": {
 53          "lr": 0.001,
 54          "weight_decay": 0.0
 55        },
 56        "mean": 0.00744940631557256
 57      },
 58      {
 59        "cfg": {
 60          "lr": 0.003,
 61          "weight_decay": 0.0
 62        },
 63        "mean": 0.002477104659192264
 64      },
 65      {
 66        "cfg": {
 67          "lr": 0.01,
 68          "weight_decay": 0.0
 69        },
 70        "mean": 0.0007909908308647573
 71      }
 72    ]
 73  },
 74  "idea": {
 75    "mean": 0.0007747690397081897,
 76    "std": 0.00018247493264257159,
 77    "per_seed": [
 78      0.0009708237485028803,
 79      0.0005817511701025069,
 80      0.0010273662628605962,
 81      0.0005840221419930458,
 82      0.0007391659892164171,
 83      0.0005464869318529963,
 84      0.0009673092281445861,
 85      0.0007812268449924886
 86    ],
 87    "n": 8
 88  },
 89  "comparison": {
 90    "delta_mean": 0.0005502363055711612,
 91    "idea_wins": 0,
 92    "n_pairs": 8,
 93    "per_seed_diffs": [
 94      0.0007529816939495504,
 95      0.00035173160722479224,
 96      0.0007518529309891164,
 97      0.00037896960566286,
 98      0.00048808843712322414,
 99      0.00030911587236914784,
100      0.0007730660581728444,
101      0.0005960842390777543
102    ],
103    "p_value": 0.0081,
104    "mde": 0.00016245867586821579,
105    "mde_rel_pct": 72.3541164243206,
106    "verdict": "idea worse (significant)",
107    "system_worked": false
108  },
109  "mechanism_signature": {
110    "selection": "best of shared 3-point LR/weight-decay grid on seeds 0-3",
111    "idea_best_cfg": {
112      "lr": 0.01,
113      "weight_decay": 0.0
114    },
115    "prediction": "nullspace decoder should make C u exactly zero while ambient output violates endpoints",
116    "predicted_boundary_residual_idea": 0.0,
117    "observed_boundary_residual_idea_max": 0.0,
118    "observed_boundary_residual_baseline_max": 0.02415449172258377,
119    "observed_boundary_residual_baseline_mean": 0.00435631675645709,
120    "observed_boundary_residual_idea_mean": 0.0,
121    "ratio_baseline_to_idea_max": 24154490880.0,
122    "confirmed": true,
123    "trained_models": true
124  },
125  "custom_track": {
126    "name": "poisson_dirichlet",
127    "file": "/home/maxwelhelp/all/math2nn/bench/custom_tracks/poisson_dirichlet.py",
128    "domain": "pde",
129    "reason": "built-in custom PDE track has fixed-grid field outputs and homogeneous Dirichlet constraints"
130  },
131  "protocol": {
132    "epochs": 20,
133    "batch": 128,
134    "seeds": [
135      0,
136      1,
137      2,
138      3,
139      4,
140      5,
141      6,
142      7
143    ],
144    "grid": [
145      {
146        "lr": 0.001,
147        "weight_decay": 0.0
148      },
149      {
150        "lr": 0.003,
151        "weight_decay": 0.0
152      },
153      {
154        "lr": 0.01,
155        "weight_decay": 0.0
156      }
157    ]
158  }
159}