Covariance-Adaptive Hermite Latent Bottleneck / bench_report.json

Mechanism confirmed, baseline not beaten

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": 0.0001
 17        },
 18        "mean": 16.534404277801514
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "weight_decay": 0.0001
 24        },
 25        "mean": 9.554428339004517
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.006,
 30          "weight_decay": 0.0001
 31        },
 32        "mean": 7.10407030582428
 33      }
 34    ],
 35    "full": {
 36      "mean": 7.061818897724152,
 37      "std": 0.3679269493562873,
 38      "per_seed": [
 39        6.67182731628418,
 40        6.9647216796875,
 41        7.721733093261719,
 42        7.057999134063721,
 43        6.471690654754639,
 44        6.978625297546387,
 45        7.271554470062256,
 46        7.3563995361328125
 47      ],
 48      "n": 8
 49    }
 50  },
 51  "idea": {
 52    "mean": 178.17990684509277,
 53    "std": 9.629824242821801,
 54    "per_seed": [
 55      190.22879028320312,
 56      175.91212463378906,
 57      179.41261291503906,
 58      165.240966796875,
 59      190.38267517089844,
 60      169.86196899414062,
 61      167.13531494140625,
 62      187.26480102539062
 63    ],
 64    "n": 8,
 65    "signatures": [
 66      {
 67        "q": 1.0,
 68        "predicted_tail": 1.0,
 69        "observed_tail": 0.33333323108538504,
 70        "degree": 0
 71      },
 72      {
 73        "q": 1.0,
 74        "predicted_tail": 1.0,
 75        "observed_tail": 0.33333318546496393,
 76        "degree": 0
 77      },
 78      {
 79        "q": 1.0,
 80        "predicted_tail": 1.0,
 81        "observed_tail": 0.3333331799108243,
 82        "degree": 0
 83      },
 84      {
 85        "q": 1.0,
 86        "predicted_tail": 1.0,
 87        "observed_tail": 0.3333332577271232,
 88        "degree": 0
 89      },
 90      {
 91        "q": 1.0,
 92        "predicted_tail": 1.0,
 93        "observed_tail": 0.3333332312105795,
 94        "degree": 0
 95      },
 96      {
 97        "q": 1.0,
 98        "predicted_tail": 1.0,
 99        "observed_tail": 0.33333320043828085,
100        "degree": 0
101      },
102      {
103        "q": 1.0,
104        "predicted_tail": 1.0,
105        "observed_tail": 0.3333332324274767,
106        "degree": 0
107      },
108      {
109        "q": 1.0,
110        "predicted_tail": 1.0,
111        "observed_tail": 0.333333177763843,
112        "degree": 0
113      }
114    ],
115    "cfg": {
116      "lr": 0.006,
117      "weight_decay": 0.0001
118    }
119  },
120  "comparison": {
121    "delta_mean": 171.11808794736862,
122    "idea_wins": 0,
123    "n_pairs": 8,
124    "per_seed_diffs": [
125      183.55696296691895,
126      168.94740295410156,
127      171.69087982177734,
128      158.18296766281128,
129      183.9109845161438,
130      162.88334369659424,
131      159.863760471344,
132      179.9084014892578
133    ],
134    "p_value": 0.0081,
135    "mde": 8.726735703488881,
136    "mde_rel_pct": 123.57631695003239,
137    "verdict": "idea worse (significant)",
138    "system_worked": false
139  },
140  "mechanism_signature": {
141    "quantity": "trained held-out latent Hermite omitted-energy proxy",
142    "predicted_mean_tail_scale": 1.0,
143    "observed_mean_tail": 0.3333332120035596,
144    "mean_q": 1.0,
145    "degrees": [
146      0,
147      0,
148      0,
149      0,
150      0,
151      0,
152      0,
153      0
154    ],
155    "confirmed": true
156  },
157  "idea_sweep": [
158    {
159      "cfg": {
160        "lr": 0.001,
161        "weight_decay": 0.0001
162      },
163      "mean": 224.16369247436523
164    },
165    {
166      "cfg": {
167        "lr": 0.003,
168        "weight_decay": 0.0001
169      },
170      "mean": 204.7395420074463
171    },
172    {
173      "cfg": {
174        "lr": 0.006,
175        "weight_decay": 0.0001
176      },
177      "mean": 178.17990684509277
178    }
179  ],
180  "protocol": {
181    "paired_seeds": [
182      0,
183      1,
184      2,
185      3,
186      4,
187      5,
188      6,
189      7
190    ],
191    "epochs": 18,
192    "batch": 128,
193    "baseline_grid": [
194      {
195        "lr": 0.001,
196        "weight_decay": 0.0001
197      },
198      {
199        "lr": 0.003,
200        "weight_decay": 0.0001
201      },
202      {
203        "lr": 0.006,
204        "weight_decay": 0.0001
205      }
206    ],
207    "idea_grid": [
208      {
209        "lr": 0.001,
210        "weight_decay": 0.0001
211      },
212      {
213        "lr": 0.003,
214        "weight_decay": 0.0001
215      },
216      {
217        "lr": 0.006,
218        "weight_decay": 0.0001
219      }
220    ],
221    "baseline_selection": "sweep_baseline on seeds 0..3, full reevaluation on 0..7"
222  }
223}