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": 0.0015
 15        },
 16        "mean": 0.0017903338593896478
 17      },
 18      {
 19        "cfg": {
 20          "lr": 0.003
 21        },
 22        "mean": 0.0011179699067724869
 23      },
 24      {
 25        "cfg": {
 26          "lr": 0.006
 27        },
 28        "mean": 0.0007884320366429165
 29      }
 30    ],
 31    "full": {
 32      "mean": 0.0007652444473933429,
 33      "std": 0.00014266442888030616,
 34      "per_seed": [
 35        0.0007294044480659068,
 36        0.0007733436650596559,
 37        0.0005735447048209608,
 38        0.0010774353286251426,
 39        0.0006340121617540717,
 40        0.0008162351441569626,
 41        0.0006983152125030756,
 42        0.0008196649141609669
 43      ],
 44      "n": 8
 45    }
 46  },
 47  "idea": {
 48    "mean": 0.0006864842143841088,
 49    "std": 0.0001532192657171232,
 50    "per_seed": [
 51      0.0007881892961449921,
 52      0.0005242026527412236,
 53      0.0005977022228762507,
 54      0.0009137507877312601,
 55      0.0007462537032552063,
 56      0.0004245340824127197,
 57      0.0006721826503053308,
 58      0.0008250583196058869
 59    ],
 60    "n": 8
 61  },
 62  "comparison": {
 63    "delta_mean": -7.876023300923407e-05,
 64    "idea_wins": 4,
 65    "n_pairs": 8,
 66    "per_seed_diffs": [
 67      5.878484807908535e-05,
 68      -0.00024914101231843233,
 69      2.4157518055289984e-05,
 70      -0.0001636845408938825,
 71      0.00011224154150113463,
 72      -0.0003917010617442429,
 73      -2.6132562197744846e-05,
 74      5.393405444920063e-06
 75    ],
 76    "p_value": 0.2712,
 77    "mde": 0.00014483215243757036,
 78    "mde_rel_pct": 18.926259828596347,
 79    "verdict": "no significant win",
 80    "system_worked": false
 81  },
 82  "mechanism_signature": {
 83    "track_structure": "controlled pendulum multi-step dynamics",
 84    "idea_sweep": [
 85      {
 86        "cfg": {
 87          "lr": 0.0015,
 88          "coeff": 0.03,
 89          "target": 0.75
 90        },
 91        "result": {
 92          "mean": 0.0016631976614007726,
 93          "std": 0.000559542889000477,
 94          "per_seed": [
 95            0.0011964859440922737,
 96            0.0018847997998818755,
 97            0.0014795587630942464,
 98            0.002758872462436557,
 99            0.0010670105693861842,
100            0.0009833034127950668,
101            0.0019550875294953585,
102            0.001980462810024619
103          ],
104          "n": 8
105        }
106      },
107      {
108        "cfg": {
109          "lr": 0.003,
110          "coeff": 0.03,
111          "target": 0.75
112        },
113        "result": {
114          "mean": 0.0010474532173248008,
115          "std": 0.0001957735370690522,
116          "per_seed": [
117            0.0008862946415320039,
118            0.0009964596247300506,
119            0.000995335285551846,
120            0.001299441559240222,
121            0.0009098018635995686,
122            0.0007802878390066326,
123            0.001124624046497047,
124            0.0013873808784410357
125          ],
126          "n": 8
127        }
128      },
129      {
130        "cfg": {
131          "lr": 0.006,
132          "coeff": 0.03,
133          "target": 0.75
134        },
135        "result": {
136          "mean": 0.0006864842143841088,
137          "std": 0.0001532192657171232,
138          "per_seed": [
139            0.0007881892961449921,
140            0.0005242026527412236,
141            0.0005977022228762507,
142            0.0009137507877312601,
143            0.0007462537032552063,
144            0.0004245340824127197,
145            0.0006721826503053308,
146            0.0008250583196058869
147          ],
148          "n": 8
149        }
150      }
151    ],
152    "predicted": "idea observed gain < baseline observed gain",
153    "baseline_mean_gain": 0.06808034889400005,
154    "idea_mean_gain": 0.06390607543289661,
155    "relative_change_pct": -6.13139258084989,
156    "confirmed": true
157  },
158  "custom_track": null
159}