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      "wbar": 0.05
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "wbar": 0.05
 17        },
 18        "mean": 0.004298779065720737
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.001,
 23          "wbar": 0.15
 24        },
 25        "mean": 0.004298779065720737
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.001,
 30          "wbar": 0.3
 31        },
 32        "mean": 0.004298779065720737
 33      },
 34      {
 35        "cfg": {
 36          "lr": 0.003,
 37          "wbar": 0.05
 38        },
 39        "mean": 0.0034615940821822733
 40      },
 41      {
 42        "cfg": {
 43          "lr": 0.003,
 44          "wbar": 0.15
 45        },
 46        "mean": 0.0034615940821822733
 47      },
 48      {
 49        "cfg": {
 50          "lr": 0.003,
 51          "wbar": 0.3
 52        },
 53        "mean": 0.0034615940821822733
 54      },
 55      {
 56        "cfg": {
 57          "lr": 0.01,
 58          "wbar": 0.05
 59        },
 60        "mean": 0.0014865802077110857
 61      },
 62      {
 63        "cfg": {
 64          "lr": 0.01,
 65          "wbar": 0.15
 66        },
 67        "mean": 0.0014865802077110857
 68      },
 69      {
 70        "cfg": {
 71          "lr": 0.01,
 72          "wbar": 0.3
 73        },
 74        "mean": 0.0014865802077110857
 75      }
 76    ],
 77    "full": {
 78      "mean": 0.0014198302378645167,
 79      "std": 0.00039839114901047844,
 80      "per_seed": [
 81        0.0009327387670055032,
 82        0.0018641706556081772,
 83        0.0010785290505737066,
 84        0.0020708823576569557,
 85        0.0009260809747502208,
 86        0.0014055331703275442,
 87        0.0016443432541564107,
 88        0.001436363672837615
 89      ],
 90      "n": 8
 91    }
 92  },
 93  "idea": {
 94    "mean": 0.005354407912818715,
 95    "std": 0.0011548517240013176,
 96    "per_seed": [
 97      0.0045272172428667545,
 98      0.004973915405571461,
 99      0.006366239860653877,
100      0.007458276581019163,
101      0.005999961402267218,
102      0.0034331066999584436,
103      0.005241193808615208,
104      0.004835352301597595
105    ],
106    "n": 8
107  },
108  "comparison": {
109    "delta_mean": 0.003934577674954198,
110    "idea_wins": 0,
111    "n_pairs": 8,
112    "per_seed_diffs": [
113      0.0035944784758612514,
114      0.0031097447499632835,
115      0.005287710810080171,
116      0.005387394223362207,
117      0.005073880427516997,
118      0.0020275735296308994,
119      0.003596850554458797,
120      0.00339898862875998
121    ],
122    "p_value": 0.0081,
123    "mde": 0.0010030486525361333,
124    "mde_rel_pct": 70.6456747987534,
125    "verdict": "idea worse (significant)",
126    "system_worked": false
127  },
128  "mechanism_signature": {
129    "chosen_cfg": {
130      "lr": 0.01,
131      "wbar": 0.05
132    },
133    "cV": 0.4,
134    "predicted_envelope": "Vdot <= -cV V + disturbance margin",
135    "signature": {
136      "quantity": "mean robust CLF residual versus disturbance",
137      "predicted_slope_abs_theta": 0.7278081178665161,
138      "observed_slope": 0.27696520090103144,
139      "relative_error": 0.6194529845209034,
140      "confirmed": false
141    }
142  },
143  "idea_sweep": [
144    {
145      "cfg": {
146        "lr": 0.01,
147        "wbar": 0.05
148      },
149      "result": {
150        "mean": 0.005354407912818715,
151        "std": 0.0011548517240013176,
152        "per_seed": [
153          0.0045272172428667545,
154          0.004973915405571461,
155          0.006366239860653877,
156          0.007458276581019163,
157          0.005999961402267218,
158          0.0034331066999584436,
159          0.005241193808615208,
160          0.004835352301597595
161        ],
162        "n": 8
163      }
164    },
165    {
166      "cfg": {
167        "lr": 0.001,
168        "wbar": 0.15
169      },
170      "result": {
171        "mean": 0.005414883635239676,
172        "std": 0.0011777990905441577,
173        "per_seed": [
174          0.004449672531336546,
175          0.0050073061138391495,
176          0.006440161261707544,
177          0.007577271666377783,
178          0.006071907002478838,
179          0.0035233551170676947,
180          0.005342461168766022,
181          0.004906934220343828
182        ],
183        "n": 8
184      }
185    },
186    {
187      "cfg": {
188        "lr": 0.01,
189        "wbar": 0.3
190      },
191      "result": {
192        "mean": 0.005623006261885166,
193        "std": 0.0012009004828101744,
194        "per_seed": [
195          0.004604719579219818,
196          0.005171346943825483,
197          0.006618546321988106,
198          0.007910414598882198,
199          0.006261724047362804,
200          0.003785517066717148,
201          0.005539596080780029,
202          0.005092185456305742
203        ],
204        "n": 8
205      }
206    }
207  ]
208}