Adaptive Zonotope Safety Shield / 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": 0.01,
 10      "epochs": 10
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "epochs": 10
 17        },
 18        "mean": 0.02325876720715314,
 19        "std": 0.009097045470607702,
 20        "per_seed": [
 21          0.006040765903890133,
 22          0.03010248765349388,
 23          0.02543320134282112,
 24          0.024085573852062225,
 25          0.010201698169112206,
 26          0.030222969129681587,
 27          0.031349945813417435,
 28          0.028633495792746544
 29        ],
 30        "n": 8
 31      },
 32      {
 33        "cfg": {
 34          "lr": 0.003,
 35          "epochs": 10
 36        },
 37        "mean": 0.003875804533890914,
 38        "std": 0.0019874791140290942,
 39        "per_seed": [
 40          0.0009708286379463971,
 41          0.003958100453019142,
 42          0.004776353016495705,
 43          0.005382170435041189,
 44          0.0012958660954609513,
 45          0.002271931618452072,
 46          0.006541829090565443,
 47          0.005809356924146414
 48        ],
 49        "n": 8
 50      },
 51      {
 52        "cfg": {
 53          "lr": 0.01,
 54          "epochs": 10
 55        },
 56        "mean": 0.002775895278318785,
 57        "std": 0.0017084856346187775,
 58        "per_seed": [
 59          0.0014615225372835994,
 60          0.004334237892180681,
 61          0.0012712624156847596,
 62          0.006652707699686289,
 63          0.0020110509358346462,
 64          0.0024756770581007004,
 65          0.0018914247630164027,
 66          0.0021092789247632027
 67        ],
 68        "n": 8
 69      }
 70    ],
 71    "harness_tuning": {
 72      "best_cfg": {
 73        "lr": 0.01,
 74        "epochs": 10
 75      },
 76      "sweep": [
 77        {
 78          "cfg": {
 79            "lr": 0.001,
 80            "epochs": 10
 81          },
 82          "mean": 0.02141550718806684
 83        },
 84        {
 85          "cfg": {
 86            "lr": 0.003,
 87            "epochs": 10
 88          },
 89          "mean": 0.0037718631356256083
 90        },
 91        {
 92          "cfg": {
 93            "lr": 0.01,
 94            "epochs": 10
 95          },
 96          "mean": 0.0034299326362088323
 97        }
 98      ],
 99      "full": {
100        "mean": 0.002775895278318785,
101        "std": 0.0017084856346187775,
102        "per_seed": [
103          0.0014615225372835994,
104          0.004334237892180681,
105          0.0012712624156847596,
106          0.006652707699686289,
107          0.0020110509358346462,
108          0.0024756770581007004,
109          0.0018914247630164027,
110          0.0021092789247632027
111        ],
112        "n": 8
113      }
114    },
115    "full": {
116      "mean": 0.002775895278318785,
117      "std": 0.0017084856346187775,
118      "per_seed": [
119        0.0014615225372835994,
120        0.004334237892180681,
121        0.0012712624156847596,
122        0.006652707699686289,
123        0.0020110509358346462,
124        0.0024756770581007004,
125        0.0018914247630164027,
126        0.0021092789247632027
127      ],
128      "n": 8
129    }
130  },
131  "idea": {
132    "per_seed": [
133      0.0018240498611703515,
134      0.0038136225193738937,
135      0.001709936186671257,
136      0.005960390437394381,
137      0.001420309068635106,
138      0.0025334814563393593,
139      0.002282342640683055,
140      0.0019062625942751765
141    ],
142    "mean": 0.0026812993455678225,
143    "std": 0.0014179568156557768,
144    "n": 8
145  },
146  "comparison": {
147    "delta_mean": -9.459593275096267e-05,
148    "idea_wins": 4,
149    "n_pairs": 8,
150    "per_seed_diffs": [
151      0.00036252732388675213,
152      -0.0005206153728067875,
153      0.0004386737709864974,
154      -0.0006923172622919083,
155      -0.0005907418671995401,
156      5.7804398238658905e-05,
157      0.0003909178776666522,
158      -0.00020301633048802614
159    ],
160    "p_value": 0.58435,
161    "mde": 0.00039333102068697407,
162    "mde_rel_pct": 14.16951942528589,
163    "verdict": "no significant win",
164    "system_worked": false
165  },
166  "mechanism_signature": {
167    "idea_sweep": [
168      {
169        "cfg": {
170          "lr": 0.001,
171          "epochs": 10,
172          "shield_lambda": 0.01
173        },
174        "mean": 0.022998466796707362,
175        "std": 0.009109715346858763,
176        "per_seed": [
177          0.005741507280617952,
178          0.029806293547153473,
179          0.02514277957379818,
180          0.02393505349755287,
181          0.00990633387118578,
182          0.029928065836429596,
183          0.031049689278006554,
184          0.02847801148891449
185        ],
186        "n": 8
187      },
188      {
189        "cfg": {
190          "lr": 0.003,
191          "epochs": 10,
192          "shield_lambda": 0.03
193        },
194        "mean": 0.0036644875799538568,
195        "std": 0.001954599449464149,
196        "per_seed": [
197          0.0011084122816100717,
198          0.0033393101766705513,
199          0.005045472178608179,
200          0.0046833050437271595,
201          0.0011171806836500764,
202          0.0019260895205661654,
203          0.006277866195887327,
204          0.0058182645589113235
205        ],
206        "n": 8
207      },
208      {
209        "cfg": {
210          "lr": 0.01,
211          "epochs": 10,
212          "shield_lambda": 0.06
213        },
214        "mean": 0.0026812993455678225,
215        "std": 0.0014179568156557768,
216        "per_seed": [
217          0.0018240498611703515,
218          0.0038136225193738937,
219          0.001709936186671257,
220          0.005960390437394381,
221          0.001420309068635106,
222          0.0025334814563393593,
223          0.002282342640683055,
224          0.0019062625942751765
225        ],
226        "n": 8
227      }
228    ],
229    "mechanism_signature": {
230      "prediction": "online residual zonotope contracts after burn-in while held-out residual coverage remains high",
231      "trained_test_mse": 0.0011084122816100717,
232      "observed_residual_center": 1.96993350982666e-05,
233      "observed_residual_radius_q": 3.6635516929626464,
234      "heldout_residual_coverage": 1.0,
235      "unsafe_point_prediction_rate": 0.05999999865889549,
236      "unsafe_reachable_interval_rate": 1.0,
237      "confirmed": false
238    }
239  },
240  "math_check": {
241    "cases": 1000,
242    "containment_disagreements": 0,
243    "max_sample_support_gap": 0.0
244  },
245  "protocol_note": "Eight paired seeds; baseline and idea share rnn_small, data, Adam, epochs, batch, and union learning rates. Lower test MSE is primary."
246}