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      "degree": 2
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "degree": 2
 17        },
 18        "mean": 0.003609913313994184
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "degree": 2
 24        },
 25        "mean": 0.004211013438180089
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.006,
 30          "degree": 2
 31        },
 32        "mean": 0.0017769129190128297
 33      },
 34      {
 35        "cfg": {
 36          "lr": 0.001,
 37          "degree": 3
 38        },
 39        "mean": 0.02138834842480719
 40      },
 41      {
 42        "cfg": {
 43          "lr": 0.003,
 44          "degree": 3
 45        },
 46        "mean": 0.005644599266815931
 47      },
 48      {
 49        "cfg": {
 50          "lr": 0.006,
 51          "degree": 3
 52        },
 53        "mean": 0.0025558681518305093
 54      }
 55    ],
 56    "full": {
 57      "mean": 0.0015814606886124238,
 58      "std": 0.0004392499281827401,
 59      "per_seed": [
 60        0.0010742004960775375,
 61        0.0021969457156956196,
 62        0.0015373978530988097,
 63        0.002299107611179352,
 64        0.001275557209737599,
 65        0.0012933816760778427,
 66        0.0011743184877559543,
 67        0.0018007764592766762
 68      ],
 69      "n": 8
 70    }
 71  },
 72  "idea": {
 73    "mean": 0.0021942039020359516,
 74    "std": 0.0007557853669569652,
 75    "per_seed": [
 76      0.0022852420806884766,
 77      0.002649907488375902,
 78      0.001380756264552474,
 79      0.003865908132866025,
 80      0.001558732008561492,
 81      0.0023790807463228703,
 82      0.0016722833970561624,
 83      0.0017617210978642106
 84    ],
 85    "n": 8,
 86    "sweep": [
 87      {
 88        "cfg": {
 89          "lr": 0.001,
 90          "threshold": 0.15
 91        },
 92        "mean": 0.02138834842480719
 93      },
 94      {
 95        "cfg": {
 96          "lr": 0.001,
 97          "threshold": 0.25
 98        },
 99        "mean": 0.021395483752712607
100      },
101      {
102        "cfg": {
103          "lr": 0.001,
104          "threshold": 0.4
105        },
106        "mean": 0.02141962433233857
107      },
108      {
109        "cfg": {
110          "lr": 0.003,
111          "threshold": 0.15
112        },
113        "mean": 0.005644599266815931
114      },
115      {
116        "cfg": {
117          "lr": 0.003,
118          "threshold": 0.25
119        },
120        "mean": 0.005646273668389767
121      },
122      {
123        "cfg": {
124          "lr": 0.003,
125          "threshold": 0.4
126        },
127        "mean": 0.005661312199663371
128      },
129      {
130        "cfg": {
131          "lr": 0.006,
132          "threshold": 0.15
133        },
134        "mean": 0.0025558681518305093
135      },
136      {
137        "cfg": {
138          "lr": 0.006,
139          "threshold": 0.25
140        },
141        "mean": 0.0025564142561051995
142      },
143      {
144        "cfg": {
145          "lr": 0.006,
146          "threshold": 0.4
147        },
148        "mean": 0.0025454534916207194
149      }
150    ]
151  },
152  "comparison": {
153    "delta_mean": 0.0006127432134235278,
154    "idea_wins": 2,
155    "n_pairs": 8,
156    "per_seed_diffs": [
157      0.001211041584610939,
158      0.0004529617726802826,
159      -0.0001566415885463357,
160      0.0015668005216866732,
161      0.00028317479882389307,
162      0.0010856990702450275,
163      0.0004979649093002081,
164      -3.905536141246557e-05
165    ],
166    "p_value": 0.0324,
167    "mde": 0.0005150508833037192,
168    "mde_rel_pct": 32.5680484511838,
169    "verdict": "idea worse (significant)",
170    "system_worked": false
171  },
172  "mechanism_signature": {
173    "confirmed": true,
174    "prediction": "cubic residual proxy scales as amplitude^3",
175    "observed_mean_proxy_ratio_at_1p5": 3.375,
176    "mean_rho": 0.8967330753803253,
177    "mean_active_fraction": 0.9953124970197678,
178    "track_match": "dynamics/control pendulum; trained GRU systems evaluated on standard test MSE"
179  },
180  "protocol_notes": "8 paired seeds; baseline and idea share LiftedGRU; 12 epochs, batch 128; all tried learning rates are in both grids."
181}