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": 0.001,
 10      "alpha": 0.7,
 11      "sigma": 0.035
 12    },
 13    "sweep": [
 14      {
 15        "cfg": {
 16          "lr": 0.001,
 17          "alpha": 0.7,
 18          "sigma": 0.035
 19        },
 20        "mean": 0.8283906430006027
 21      },
 22      {
 23        "cfg": {
 24          "lr": 0.003,
 25          "alpha": 0.7,
 26          "sigma": 0.035
 27        },
 28        "mean": 0.8293724805116653
 29      },
 30      {
 31        "cfg": {
 32          "lr": 0.01,
 33          "alpha": 0.7,
 34          "sigma": 0.035
 35        },
 36        "mean": 0.8328554034233093
 37      }
 38    ],
 39    "full": {
 40      "mean": 0.787501223385334,
 41      "std": 0.06788976443931306,
 42      "per_seed": [
 43        0.8156408071517944,
 44        0.7547019720077515,
 45        0.8308971524238586,
 46        0.9123226404190063,
 47        0.7518438100814819,
 48        0.7541651725769043,
 49        0.8133183717727661,
 50        0.6671198606491089
 51      ],
 52      "n": 8
 53    }
 54  },
 55  "idea": {
 56    "mean": 0.7941433116793633,
 57    "std": 0.07100903064007838,
 58    "per_seed": [
 59      0.8096513748168945,
 60      0.803184986114502,
 61      0.8362289667129517,
 62      0.9341802597045898,
 63      0.7649334669113159,
 64      0.7608115077018738,
 65      0.7781288623809814,
 66      0.6660270690917969
 67    ],
 68    "n": 8
 69  },
 70  "comparison": {
 71    "delta_mean": 0.006642088294029236,
 72    "idea_wins": 3,
 73    "n_pairs": 8,
 74    "per_seed_diffs": [
 75      -0.005989432334899902,
 76      0.04848301410675049,
 77      0.005331814289093018,
 78      0.021857619285583496,
 79      0.013089656829833984,
 80      0.006646335124969482,
 81      -0.03518950939178467,
 82      -0.0010927915573120117
 83    ],
 84    "p_value": 0.4855,
 85    "mde": 0.0199747259489214,
 86    "mde_rel_pct": 2.5364691959529213,
 87    "verdict": "no measurable effect",
 88    "system_worked": false
 89  },
 90  "mechanism_signature": {
 91    "idea_config": {
 92      "lr": 0.001,
 93      "alpha": 0.7,
 94      "sigma": 0.035
 95    },
 96    "idea_sweep": [
 97      {
 98        "cfg": {
 99          "lr": 0.001,
100          "alpha": 0.7,
101          "sigma": 0.035
102        },
103        "result": {
104          "mean": 0.7941433116793633,
105          "std": 0.07100903064007838,
106          "per_seed": [
107            0.8096513748168945,
108            0.803184986114502,
109            0.8362289667129517,
110            0.9341802597045898,
111            0.7649334669113159,
112            0.7608115077018738,
113            0.7781288623809814,
114            0.6660270690917969
115          ],
116          "n": 8
117        }
118      },
119      {
120        "cfg": {
121          "lr": 0.003,
122          "alpha": 0.7,
123          "sigma": 0.035
124        },
125        "result": {
126          "mean": 0.8040007650852203,
127          "std": 0.07647255737581943,
128          "per_seed": [
129            0.8088628649711609,
130            0.8104173541069031,
131            0.8454757928848267,
132            0.9598079919815063,
133            0.7718257308006287,
134            0.7834799289703369,
135            0.7857349514961243,
136            0.6664015054702759
137          ],
138          "n": 8
139        }
140      },
141      {
142        "cfg": {
143          "lr": 0.01,
144          "alpha": 0.7,
145          "sigma": 0.035
146        },
147        "result": {
148          "mean": 0.8066116720438004,
149          "std": 0.07406785573886294,
150          "per_seed": [
151            0.811928391456604,
152            0.8005879521369934,
153            0.8522548079490662,
154            0.9512794017791748,
155            0.7587114572525024,
156            0.7952184677124023,
157            0.8118025064468384,
158            0.6711103916168213
159          ],
160          "n": 8
161        }
162      }
163    ],
164    "sampler": "feedback pendulum rollout + Gaussian exploration + uniform reservoir",
165    "mechanism_signature": {
166      "prediction": "gradient sign predicts which bounded control increases model-predicted next angle",
167      "observed_sign_agreement": 1.0,
168      "predicted_threshold": 0.75,
169      "confirmed": true
170    },
171    "protocol_note": "8 paired seeds; baseline and idea share rnn_small, Adam, epochs, batch, and lr union."
172  }
173}