1{
  2  "bench_version": 1,
  3  "track": "tabular",
  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": 1,
 11      "wd": 0.0001
 12    },
 13    "sweep": [
 14      {
 15        "cfg": {
 16          "lr": 0.001,
 17          "epochs": 1,
 18          "wd": 0.0
 19        },
 20        "mean": 236.9449577331543
 21      },
 22      {
 23        "cfg": {
 24          "lr": 0.001,
 25          "epochs": 1,
 26          "wd": 0.0001
 27        },
 28        "mean": 236.94502639770508
 29      },
 30      {
 31        "cfg": {
 32          "lr": 0.003,
 33          "epochs": 1,
 34          "wd": 0.0
 35        },
 36        "mean": 235.13161849975586
 37      },
 38      {
 39        "cfg": {
 40          "lr": 0.003,
 41          "epochs": 1,
 42          "wd": 0.0001
 43        },
 44        "mean": 235.1318359375
 45      },
 46      {
 47        "cfg": {
 48          "lr": 0.01,
 49          "epochs": 1,
 50          "wd": 0.0
 51        },
 52        "mean": 228.67240142822266
 53      },
 54      {
 55        "cfg": {
 56          "lr": 0.01,
 57          "epochs": 1,
 58          "wd": 0.0001
 59        },
 60        "mean": 228.67166900634766
 61      }
 62    ],
 63    "full": {
 64      "mean": 231.52675819396973,
 65      "std": 21.300660878391202,
 66      "per_seed": [
 67        256.47283935546875,
 68        229.34703063964844,
 69        246.06045532226562,
 70        182.8063507080078,
 71        247.07217407226562,
 72        224.0729522705078,
 73        240.7747039794922,
 74        225.60755920410156
 75      ],
 76      "n": 8
 77    }
 78  },
 79  "idea": {
 80    "mean": 489.4525787830353,
 81    "std": 991.4184272998954,
 82    "per_seed": [
 83      29.48086166381836,
 84      76.03532409667969,
 85      67.20081329345703,
 86      22.497228622436523,
 87      3082.7841796875,
 88      505.4776916503906,
 89      78.81880950927734,
 90      53.325721740722656
 91    ],
 92    "n": 8
 93  },
 94  "comparison": {
 95    "delta_mean": 257.92582058906555,
 96    "idea_wins": 6,
 97    "n_pairs": 8,
 98    "per_seed_diffs": [
 99      -226.9919776916504,
100      -153.31170654296875,
101      -178.8596420288086,
102      -160.3091220855713,
103      2835.7120056152344,
104      281.4047393798828,
105      -161.95589447021484,
106      -172.2818374633789
107    ],
108    "p_value": 0.9404,
109    "mde": 881.3323499422643,
110    "mde_rel_pct": 380.66111961188386,
111    "verdict": "no significant win",
112    "system_worked": false
113  },
114  "mechanism_signature": {
115    "prediction": "larger cross-block certificate predicts larger observed projected Fisher spectral distortion",
116    "predicted_values": [
117      1728.5964137448777,
118      1842.2349990275347,
119      580.0484882837169,
120      667.8731875822858,
121      3872.9491724735503,
122      1046.0315549770417,
123      1335.0321904102336,
124      1300.5161090698366
125    ],
126    "observed_values": [
127      0.0,
128      0.0,
129      0.0,
130      0.0,
131      0.0,
132      0.0,
133      0.0,
134      0.0
135    ],
136    "correlation": 0.0,
137    "confirmed": false
138  },
139  "idea_sweep": [
140    {
141      "cfg": {
142        "lr": 0.001,
143        "epochs": 1,
144        "block": 64,
145        "tau": 0.02,
146        "damp": 0.01,
147        "sketch": 3
148      },
149      "mean": 228.61152267456055
150    },
151    {
152      "cfg": {
153        "lr": 0.003,
154        "epochs": 1,
155        "block": 64,
156        "tau": 0.02,
157        "damp": 0.01,
158        "sketch": 3
159      },
160      "mean": 200.8146095275879
161    },
162    {
163      "cfg": {
164        "lr": 0.01,
165        "epochs": 1,
166        "block": 64,
167        "tau": 0.02,
168        "damp": 0.01,
169        "sketch": 3
170      },
171      "mean": 48.8035569190979
172    }
173  ],
174  "chosen_idea_cfg": {
175    "lr": 0.01,
176    "epochs": 1,
177    "block": 64,
178    "tau": 0.02,
179    "damp": 0.01,
180    "sketch": 3
181  },
182  "protocol_notes": "Tabular matches an optimizer/preconditioning intervention; both systems use identical mlp_tiny and paired seeds. Fisher uses a small per-example gradient sketch and block-local factors."
183}