Residual-to-State Update Throttle / bench_report.json

Failed on benchmark

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": 0.01,
 10      "weight_decay": 0.0
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "weight_decay": 0.0
 17        },
 18        "mean": 0.0016094102757051587
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "weight_decay": 0.0
 24        },
 25        "mean": 0.0009143295173998922
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.01,
 30          "weight_decay": 0.0
 31        },
 32        "mean": 0.0004333868491812609
 33      }
 34    ],
 35    "full": {
 36      "mean": 0.00040872101453715004,
 37      "std": 7.309234212470366e-05,
 38      "per_seed": [
 39        0.0003514306154102087,
 40        0.00042900562402792275,
 41        0.0004498512134887278,
 42        0.0005032599437981844,
 43        0.00025691225891932845,
 44        0.0003992249839939177,
 45        0.00039930528146214783,
 46        0.0004807781951967627
 47      ],
 48      "n": 8
 49    },
 50    "grid_union": [
 51      {
 52        "lr": 0.001,
 53        "weight_decay": 0.0
 54      },
 55      {
 56        "lr": 0.003,
 57        "weight_decay": 0.0
 58      },
 59      {
 60        "lr": 0.01,
 61        "weight_decay": 0.0
 62      }
 63    ]
 64  },
 65  "idea": {
 66    "mean": 0.000397826104745036,
 67    "std": 0.00010231631988854813,
 68    "per_seed": [
 69      0.0002917043457273394,
 70      0.0005510751507245004,
 71      0.00044953575707040727,
 72      0.00048635262646712363,
 73      0.00025691225891932845,
 74      0.0002804143587127328,
 75      0.00039916730020195246,
 76      0.0004674470401369035
 77    ],
 78    "n": 8,
 79    "sweep_grid": [
 80      {
 81        "lr": 0.001,
 82        "weight_decay": 0.0,
 83        "kappa": 0.5
 84      },
 85      {
 86        "lr": 0.003,
 87        "weight_decay": 0.0,
 88        "kappa": 0.5
 89      },
 90      {
 91        "lr": 0.01,
 92        "weight_decay": 0.0,
 93        "kappa": 0.5
 94      }
 95    ]
 96  },
 97  "comparison": {
 98    "delta_mean": -1.089490979211405e-05,
 99    "idea_wins": 6,
100    "n_pairs": 8,
101    "per_seed_diffs": [
102      -5.9726269682869315e-05,
103      0.00012206952669657767,
104      -3.154564183205366e-07,
105      -1.6907317331060767e-05,
106      0.0,
107      -0.00011881062528118491,
108      -1.379812601953745e-07,
109      -1.3331155059859157e-05
110    ],
111    "p_value": 0.58435,
112    "mde": 5.6621048472105967e-05,
113    "mde_rel_pct": 13.853226640725977,
114    "verdict": "no significant win",
115    "system_worked": false
116  },
117  "mechanism_signature": {
118    "track_match": "stability/control -> actuated pendulum dynamics",
119    "idea_cfg": {
120      "lr": 0.01,
121      "weight_decay": 0.0,
122      "kappa": 0.5
123    },
124    "mechanism_signature": {
125      "residual_norm": 0.09767141938209534,
126      "sqrt_q": 14.045147895812988,
127      "delta": 0.006954104173928499,
128      "gain": 1.0,
129      "kappa": 0.5,
130      "prediction": "delta>kappa implies a=kappa/(delta+eps) and normalized residual forcing is capped",
131      "observed_throttled_fraction": 0.023809523809523808,
132      "confirmed": false
133    }
134  }
135}