Composed Trusted Reachable Families for Recurrent Networks / 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    },
 11    "sweep": [
 12      {
 13        "cfg": {
 14          "lr": 0.001
 15        },
 16        "mean": 0.004298779065720737
 17      },
 18      {
 19        "cfg": {
 20          "lr": 0.003
 21        },
 22        "mean": 0.0034615940821822733
 23      },
 24      {
 25        "cfg": {
 26          "lr": 0.006
 27        },
 28        "mean": 0.0018282315868418664
 29      }
 30    ],
 31    "full": {
 32      "mean": 0.002105819425196387,
 33      "std": 0.000389436566226016,
 34      "per_seed": [
 35        0.002004395006224513,
 36        0.001437882543541491,
 37        0.0015748648438602686,
 38        0.002295783953741193,
 39        0.0024020641576498747,
 40        0.0024898042902350426,
 41        0.0020851334556937218,
 42        0.0025566271506249905
 43      ],
 44      "n": 8
 45    }
 46  },
 47  "idea": {
 48    "mean": 0.001743134795106016,
 49    "std": 0.0006819343183841201,
 50    "per_seed": [
 51      0.002026122761890292,
 52      0.0009872562950477004,
 53      0.00172160635702312,
 54      0.003036356531083584,
 55      0.0009595885640010238,
 56      0.0010313111124560237,
 57      0.002219313522800803,
 58      0.001963523216545582
 59    ],
 60    "n": 8
 61  },
 62  "comparison": {
 63    "delta_mean": -0.0003626846300903708,
 64    "idea_wins": 4,
 65    "n_pairs": 8,
 66    "per_seed_diffs": [
 67      2.1727755665779114e-05,
 68      -0.0004506262484937906,
 69      0.00014674151316285133,
 70      0.000740572577342391,
 71      -0.001442475593648851,
 72      -0.0014584931777790189,
 73      0.0001341800671070814,
 74      -0.0005931039340794086
 75    ],
 76    "p_value": 0.2468,
 77    "mde": 0.0006552640413442428,
 78    "mde_rel_pct": 31.116820060824228,
 79    "verdict": "no significant win",
 80    "system_worked": false
 81  },
 82  "mechanism_signature": {
 83    "method": "multi_step_affine_reachability_penalty",
 84    "alpha": 0.02,
 85    "signature": {
 86      "radii": [
 87        0.02,
 88        0.04,
 89        0.08
 90      ],
 91      "observed_violation": [
 92        8.657040780235548e-06,
 93        3.4265096473973244e-05,
 94        0.00013634731294587255
 95      ],
 96      "observed_log_slope": 1.9886342514929614,
 97      "predicted_slope": 2.0,
 98      "confirmed": true
 99    },
100    "idea_lr_results": {
101      "0.001": {
102        "mean": 0.005559288372751325,
103        "std": 0.002875018990704497,
104        "per_seed": [
105          0.0068917470052838326,
106          0.0023556745145469904,
107          0.0028747664764523506,
108          0.003830716712400317,
109          0.0052271741442382336,
110          0.003931319806724787,
111          0.007791006471961737,
112          0.011571901850402355
113        ],
114        "n": 8
115      },
116      "0.003": {
117        "mean": 0.0027528391947271302,
118        "std": 0.0013131710105052272,
119        "per_seed": [
120          0.0018411790952086449,
121          0.0037763440050184727,
122          0.0017155342502519488,
123          0.005654776003211737,
124          0.0015558436280116439,
125          0.0030924330931156874,
126          0.002543396782130003,
127          0.0018432067008689046
128        ],
129        "n": 8
130      },
131      "0.006": {
132        "mean": 0.001743134795106016,
133        "std": 0.0006819343183841201,
134        "per_seed": [
135          0.002026122761890292,
136          0.0009872562950477004,
137          0.00172160635702312,
138          0.003036356531083584,
139          0.0009595885640010238,
140          0.0010313111124560237,
141          0.002219313522800803,
142          0.001963523216545582
143        ],
144        "n": 8
145      }
146    }
147  },
148  "protocol_notes": {
149    "structural_match": "dynamics recurrent/control track",
150    "same_architecture": true,
151    "baseline_lr_union": [
152      0.001,
153      0.003,
154      0.006
155    ],
156    "idea_lr_union": [
157      0.001,
158      0.003,
159      0.006
160    ],
161    "epochs": 12,
162    "train_samples": 400,
163    "test_samples": 200
164  }
165}