Truncated Volterra Stabilizer for Recurrent Blocks / 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      "weight_decay": 0.0
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "weight_decay": 0.0
 17        },
 18        "mean": 0.0012006517936242744
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "weight_decay": 0.0
 24        },
 25        "mean": 0.0006976972363190725
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.006,
 30          "weight_decay": 0.0
 31        },
 32        "mean": 0.00046024548646528274
 33      }
 34    ],
 35    "full": {
 36      "mean": 0.0003940512724511791,
 37      "std": 0.00011379624857509869,
 38      "per_seed": [
 39        0.00040922293555922806,
 40        0.0005125142051838338,
 41        0.0002987775078509003,
 42        0.0006204672972671688,
 43        0.00030420164694078267,
 44        0.00036726761027239263,
 45        0.0002520524722058326,
 46        0.00038790650432929397
 47      ],
 48      "n": 8
 49    }
 50  },
 51  "idea": {
 52    "mean": 0.00023149568187363911,
 53    "std": 6.189006390488855e-05,
 54    "per_seed": [
 55      0.0002818358188960701,
 56      0.00030632453854195774,
 57      0.00019047960813622922,
 58      0.00032223673770204186,
 59      0.00021427450701594353,
 60      0.00020972064521629363,
 61      0.00012757130025420338,
 62      0.00019952229922637343
 63    ],
 64    "n": 8
 65  },
 66  "comparison": {
 67    "delta_mean": -0.00016255559057754,
 68    "idea_wins": 8,
 69    "n_pairs": 8,
 70    "per_seed_diffs": [
 71      -0.00012738711666315794,
 72      -0.0002061896666418761,
 73      -0.00010829789971467108,
 74      -0.0002982305595651269,
 75      -8.992713992483914e-05,
 76      -0.000157546965056099,
 77      -0.00012448117195162922,
 78      -0.00018838420510292053
 79    ],
 80    "p_value": 0.0081,
 81    "mde": 5.645396435641711e-05,
 82    "mde_rel_pct": 14.326552990235925,
 83    "verdict": "idea better (significant)",
 84    "system_worked": true
 85  },
 86  "math_check": {
 87    "ordered_pairs_L8": 36,
 88    "remainder_over_quadratic_loglog_slope": 1.0000000000000004,
 89    "stable_linear_radius": true
 90  },
 91  "idea_sweep": [
 92    {
 93      "cfg": {
 94        "lr": 0.001,
 95        "weight_decay": 0.0
 96      },
 97      "mean": 0.0008330114796990529
 98    },
 99    {
100      "cfg": {
101        "lr": 0.003,
102        "weight_decay": 0.0
103      },
104      "mean": 0.00044643200089922175
105    },
106    {
107      "cfg": {
108        "lr": 0.006,
109        "weight_decay": 0.0
110      },
111      "mean": 0.00027521917581907474
112    }
113  ],
114  "idea_best_cfg": {
115    "lr": 0.006,
116    "weight_decay": 0.0
117  },
118  "parameter_counts": {
119    "baseline": 13313,
120    "idea": 15617
121  },
122  "custom_track": null,
123  "mechanism_signature": {
124    "predicted_power": 2.0,
125    "observed_power_from_trained_model": 1.0769089683295592,
126    "scale_ratio_observed": 2.1095115232492603,
127    "hidden_feature_correlation": 0.9850645661354065,
128    "test_metric_for_signature_run": 0.0002818358188960701,
129    "confirmed": false
130  }
131}