Spectral-Gated Parallel Best Responses / bench_report.json

Failed on benchmark

Raw ⬇ ZIP
  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": 0.006,
 10      "weight_decay": 0.0001
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "weight_decay": 0.0
 17        },
 18        "mean": 25.60164165496826
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.001,
 23          "weight_decay": 0.0001
 24        },
 25        "mean": 25.508389472961426
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.003,
 30          "weight_decay": 0.0
 31        },
 32        "mean": 13.860755920410156
 33      },
 34      {
 35        "cfg": {
 36          "lr": 0.003,
 37          "weight_decay": 0.0001
 38        },
 39        "mean": 13.839483976364136
 40      },
 41      {
 42        "cfg": {
 43          "lr": 0.006,
 44          "weight_decay": 0.0
 45        },
 46        "mean": 10.442761182785034
 47      },
 48      {
 49        "cfg": {
 50          "lr": 0.006,
 51          "weight_decay": 0.0001
 52        },
 53        "mean": 10.238675355911255
 54      }
 55    ],
 56    "full": {
 57      "mean": 10.678602695465088,
 58      "std": 1.2286671921700971,
 59      "per_seed": [
 60        10.056567192077637,
 61        11.58691120147705,
 62        9.562552452087402,
 63        9.74867057800293,
 64        8.789999961853027,
 65        11.738417625427246,
 66        11.37514591217041,
 67        12.570556640625
 68      ],
 69      "n": 8
 70    }
 71  },
 72  "idea": {
 73    "cfg": {
 74      "lr": 0.006,
 75      "reg": 0.02,
 76      "alpha": 0.5
 77    },
 78    "mean": 98.17565250396729,
 79    "std": 15.026422279731827,
 80    "per_seed": [
 81      108.2091293334961,
 82      109.38567352294922,
 83      88.7021255493164,
 84      95.20142364501953,
 85      67.35282135009766,
 86      112.1347427368164,
 87      89.88146209716797,
 88      114.537841796875
 89    ],
 90    "n": 8
 91  },
 92  "comparison": {
 93    "delta_mean": 87.4970498085022,
 94    "idea_wins": 0,
 95    "n_pairs": 8,
 96    "per_seed_diffs": [
 97      98.15256214141846,
 98      97.79876232147217,
 99      79.139573097229,
100      85.4527530670166,
101      58.56282138824463,
102      100.39632511138916,
103      78.50631618499756,
104      101.96728515625
105    ],
106    "p_value": 0.0081,
107    "mde": 12.591954120737624,
108    "mde_rel_pct": 117.91761974705814,
109    "verdict": "idea worse (significant)",
110    "system_worked": false
111  },
112  "mechanism_signature": {
113    "predicted_vs_observed": [
114      {
115        "seed": 0,
116        "rho_pred": 0.6043921917676925,
117        "observed_block_response": [
118          37.716896057128906,
119          54.9484977722168,
120          154.05938720703125
121        ]
122      },
123      {
124        "seed": 1,
125        "rho_pred": 0.5927385747432707,
126        "observed_block_response": [
127          39.42379379272461,
128          57.7188835144043,
129          155.12850952148438
130        ]
131      },
132      {
133        "seed": 2,
134        "rho_pred": 0.5786523669958115,
135        "observed_block_response": [
136          42.31975173950195,
137          50.240013122558594,
138          138.11692810058594
139        ]
140      },
141      {
142        "seed": 3,
143        "rho_pred": 0.560906821489334,
144        "observed_block_response": [
145          43.896156311035156,
146          62.861045837402344,
147          148.82882690429688
148        ]
149      },
150      {
151        "seed": 4,
152        "rho_pred": 0.5791306406259537,
153        "observed_block_response": [
154          45.193443298339844,
155          60.726924896240234,
156          165.0826416015625
157        ]
158      },
159      {
160        "seed": 5,
161        "rho_pred": 0.5629717469215392,
162        "observed_block_response": [
163          45.11399841308594,
164          62.12171173095703,
165          140.21766662597656
166        ]
167      },
168      {
169        "seed": 6,
170        "rho_pred": 0.5911920607089995,
171        "observed_block_response": [
172          41.594322204589844,
173          59.584537506103516,
174          163.77099609375
175        ]
176      },
177      {
178        "seed": 7,
179        "rho_pred": 0.5964572012424468,
180        "observed_block_response": [
181          46.23796081542969,
182          56.21791076660156,
183          158.2962188720703
184        ]
185      }
186    ],
187    "prediction": "gated blocks should avoid unstable coupling",
188    "confirmed": false,
189    "note": "NN probe is heuristic, not exact Hessian spectral radius"
190  },
191  "idea_sweep": [
192    {
193      "cfg": {
194        "lr": 0.001,
195        "reg": 0.02,
196        "alpha": 0.5
197      },
198      "mean": 227.52948379516602,
199      "std": 6.764206456146394,
200      "per_seed": [
201        234.45175170898438,
202        228.4759063720703,
203        232.97796630859375,
204        214.48780822753906,
205        232.79397583007812,
206        225.4188232421875,
207        219.51397705078125,
208        232.11566162109375
209      ],
210      "n": 8
211    },
212    {
213      "cfg": {
214        "lr": 0.003,
215        "reg": 0.02,
216        "alpha": 0.5
217      },
218      "mean": 198.1618824005127,
219      "std": 7.105863534863813,
220      "per_seed": [
221        205.67877197265625,
222        202.6162872314453,
223        200.92047119140625,
224        187.7106170654297,
225        191.00547790527344,
226        201.2646484375,
227        189.39215087890625,
228        206.70663452148438
229      ],
230      "n": 8
231    },
232    {
233      "cfg": {
234        "lr": 0.006,
235        "reg": 0.02,
236        "alpha": 0.5
237      },
238      "mean": 98.17565250396729,
239      "std": 15.026422279731827,
240      "per_seed": [
241        108.2091293334961,
242        109.38567352294922,
243        88.7021255493164,
244        95.20142364501953,
245        67.35282135009766,
246        112.1347427368164,
247        89.88146209716797,
248        114.537841796875
249      ],
250      "n": 8
251    }
252  ],
253  "custom_track": null
254}