Cyclic Lie-Bracket Residual Block / bench_report.json

✓✓ Beats tuned baseline

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": 0.006,
 10      "step": 0.0
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "step": 0.0
 17        },
 18        "mean": 0.0013661633856827393
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "step": 0.0
 24        },
 25        "mean": 0.0007308490021387115
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.006,
 30          "step": 0.0
 31        },
 32        "mean": 0.00039502444451500196
 33      }
 34    ],
 35    "full": {
 36      "cfg": {
 37        "lr": 0.006,
 38        "step": 0.0
 39      },
 40      "mean": 0.00039502444451500196,
 41      "std": 0.00012923294272245648,
 42      "per_seed": [
 43        0.0004081559309270233,
 44        0.0005658670561388135,
 45        0.00031052311533130705,
 46        0.0006247127312235534,
 47        0.00027073538512922823,
 48        0.00039440052933059633,
 49        0.00022947481193114072,
 50        0.00035632599610835314
 51      ],
 52      "n": 8
 53    }
 54  },
 55  "idea": {
 56    "cfg": {
 57      "lr": 0.006,
 58      "step": 0.1
 59    },
 60    "mean": 0.00018664514300326118,
 61    "std": 6.30393270484372e-05,
 62    "per_seed": [
 63      0.0002500905538909137,
 64      0.00021891914366278797,
 65      0.0001631042396184057,
 66      0.0002553996164351702,
 67      9.480195876676589e-05,
 68      0.00010218997340416536,
 69      0.00015309224545489997,
 70      0.00025556341279298067
 71    ],
 72    "n": 8,
 73    "best_cfg": {
 74      "lr": 0.006,
 75      "step": 0.1
 76    },
 77    "sweep": [
 78      {
 79        "cfg": {
 80          "lr": 0.006,
 81          "step": 0.1
 82        },
 83        "mean": 0.00018664514300326118
 84      },
 85      {
 86        "cfg": {
 87          "lr": 0.006,
 88          "step": 0.15
 89        },
 90        "mean": 0.0001907046962514869
 91      },
 92      {
 93        "cfg": {
 94          "lr": 0.006,
 95          "step": 0.22
 96        },
 97        "mean": 0.00020175300687697018
 98      }
 99    ]
100  },
101  "comparison": {
102    "delta_mean": -0.00020837930151174078,
103    "idea_wins": 8,
104    "n_pairs": 8,
105    "per_seed_diffs": [
106      -0.00015806537703610957,
107      -0.0003469479124760255,
108      -0.00014741887571290135,
109      -0.00036931311478838325,
110      -0.00017593342636246234,
111      -0.00029221055592643097,
112      -7.638256647624075e-05,
113      -0.00010076258331537247
114    ],
115    "p_value": 0.0081,
116    "mde": 9.39821994106377e-05,
117    "mde_rel_pct": 23.791489543394196,
118    "verdict": "idea better (significant)",
119    "system_worked": true
120  },
121  "mechanism_signature": {
122    "order_gap_mean": 0.012205412611365318,
123    "steps": [
124      0.009999999776482582,
125      0.019999999552965164,
126      0.03999999910593033
127    ],
128    "gaps": [
129      1.3433276762953028e-05,
130      0.00010710919741541147,
131      0.0008460786193609238
132    ],
133    "predicted_exponent": 2.0,
134    "observed_exponent": 2.9884542718353977,
135    "confirmed": false
136  }
137}