Complete Interval Abstraction Training / bench_report.json

Failed on benchmark

Raw ⬇ ZIP
  1{
  2  "bench_version": 1,
  3  "track": "dynamics",
  4  "model": "rnn_small",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.006,
 10      "epochs": 10
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "epochs": 10
 17        },
 18        "mean": 0.02098797250073403
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "epochs": 10
 24        },
 25        "mean": 0.003764481545658782
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.006,
 30          "epochs": 10
 31        },
 32        "mean": 0.002816903230268508
 33      }
 34    ],
 35    "full": {
 36      "mean": 0.0035593539651017636,
 37      "std": 0.001430033924501602,
 38      "per_seed": [
 39        0.003789895214140415,
 40        0.0012226301478222013,
 41        0.004499303642660379,
 42        0.001755783916451037,
 43        0.005420389119535685,
 44        0.0051945955492556095,
 45        0.0037620163056999445,
 46        0.0028302178252488375
 47      ],
 48      "n": 8
 49    }
 50  },
 51  "idea": {
 52    "mean": 0.0036058283294551075,
 53    "std": 0.0012201842547257958,
 54    "per_seed": [
 55      0.004097327124327421,
 56      0.0013652343768626451,
 57      0.0035981065593659878,
 58      0.0025071303825825453,
 59      0.0055099669843912125,
 60      0.0048635597340762615,
 61      0.0037585990503430367,
 62      0.0031467024236917496
 63    ],
 64    "n": 8
 65  },
 66  "comparison": {
 67    "delta_mean": 4.6474364353343844e-05,
 68    "idea_wins": 3,
 69    "n_pairs": 8,
 70    "per_seed_diffs": [
 71      0.000307431910187006,
 72      0.0001426042290404439,
 73      -0.0009011970832943916,
 74      0.0007513464661315084,
 75      8.957786485552788e-05,
 76      -0.000331035815179348,
 77      -3.4172553569078445e-06,
 78      0.0003164845984429121
 79    ],
 80    "p_value": 0.81695,
 81    "mde": 0.00041109428784617027,
 82    "mde_rel_pct": 11.549688282671736,
 83    "verdict": "no significant win",
 84    "system_worked": false
 85  },
 86  "sanity_check": {
 87    "containment": [
 88      1.0,
 89      1.0,
 90      1.0,
 91      1.0
 92    ],
 93    "h": [
 94      0.2,
 95      0.1,
 96      0.05,
 97      0.025
 98    ],
 99    "mean_width": [
100      0.2894333333333333,
101      0.18980000000000002,
102      0.14003333333333334,
103      0.11466666666666667
104    ],
105    "all_contained": true,
106    "width_decreases": true
107  },
108  "idea_config_sweep": [
109    {
110      "cfg": {
111        "lr": 0.006,
112        "epochs": 10
113      },
114      "result": {
115        "mean": 0.0036058283294551075,
116        "std": 0.0012201842547257958,
117        "per_seed": [
118          0.004097327124327421,
119          0.0013652343768626451,
120          0.0035981065593659878,
121          0.0025071303825825453,
122          0.0055099669843912125,
123          0.0048635597340762615,
124          0.0037585990503430367,
125          0.0031467024236917496
126        ],
127        "n": 8
128      }
129    },
130    {
131      "cfg": {
132        "lr": 0.001,
133        "epochs": 10
134      },
135      "result": {
136        "mean": 0.023580193694215268,
137        "std": 0.009230619863742425,
138        "per_seed": [
139          0.0067050703801214695,
140          0.030281851068139076,
141          0.026058034971356392,
142          0.023178257048130035,
143          0.010180371813476086,
144          0.03046376071870327,
145          0.03332953527569771,
146          0.028444668278098106
147        ],
148        "n": 8
149      }
150    },
151    {
152      "cfg": {
153        "lr": 0.003,
154        "epochs": 10
155      },
156      "result": {
157        "mean": 0.004334091427153908,
158        "std": 0.0020653212924998296,
159        "per_seed": [
160          0.0013217671075835824,
161          0.005213679280132055,
162          0.0035173860378563404,
163          0.0064000352285802364,
164          0.0018205104861408472,
165          0.003039174247533083,
166          0.007091913837939501,
167          0.006268265191465616
168        ],
169        "n": 8
170      }
171    }
172  ],
173  "protocol_note": "Baseline and idea share rnn_small, dataset, lr/epoch union, batch and seeds; idea differs only by interval training loss.",
174  "mechanism_signature": {
175    "predicted": "smaller h should reduce interval width while residual inflation preserves containment",
176    "observed_containment_mean": 0.7031250298023224,
177    "observed_width_mean": 0.1384880207479,
178    "observed_violation_mean": 0.010517213333514519,
179    "confirmed": false
180  },
181  "custom_track": null
182}