Bellman-Resolvent Uncertainty Targets / bench_results.json

Failed on benchmark

Raw ⬇ ZIP
  1{
  2  "bench_report": {
  3    "bench_version": 1,
  4    "track": "dynamics",
  5    "model": "rnn_small",
  6    "metric_direction": "lower is better",
  7    "n_seeds": 8,
  8    "baseline": {
  9      "best_cfg": {
 10        "lr": 0.01
 11      },
 12      "sweep": [
 13        {
 14          "cfg": {
 15            "lr": 0.001
 16          },
 17          "mean": 0.005440622277092189
 18        },
 19        {
 20          "cfg": {
 21            "lr": 0.003
 22          },
 23          "mean": 0.0012262430100236088
 24        },
 25        {
 26          "cfg": {
 27            "lr": 0.01
 28          },
 29          "mean": 0.000702349963830784
 30        }
 31      ],
 32      "full": {
 33        "mean": 0.0006528829871967901,
 34        "std": 0.0001692890895210604,
 35        "per_seed": [
 36          0.0006883637397550046,
 37          0.00054119125707075,
 38          0.0005475711077451706,
 39          0.0010322737507522106,
 40          0.00068665825529024,
 41          0.00041018708725459874,
 42          0.0006441730074584484,
 43          0.0006726456922478974
 44        ],
 45        "n": 8
 46      }
 47    },
 48    "idea": {
 49      "mean": 0.0007300472607312258,
 50      "std": 0.0002903965020952602,
 51      "per_seed": [
 52        0.0010002460330724716,
 53        0.00031693512573838234,
 54        0.0007908044499345124,
 55        0.0009665557299740613,
 56        0.000615738972555846,
 57        0.000473311694804579,
 58        0.00047172754420898855,
 59        0.0012050585355609655
 60      ],
 61      "n": 8
 62    },
 63    "comparison": {
 64      "delta_mean": 7.716427353443578e-05,
 65      "idea_wins": 4,
 66      "n_pairs": 8,
 67      "per_seed_diffs": [
 68        0.000311882293317467,
 69        -0.00022425613133236766,
 70        0.00024323334218934178,
 71        -6.571802077814937e-05,
 72        -7.091928273439407e-05,
 73        6.312460754998028e-05,
 74        -0.00017244546324945986,
 75        0.0005324128433130682
 76      ],
 77      "p_value": 0.43375,
 78      "mde": 0.00022023526291826335,
 79      "mde_rel_pct": 33.732731168851956,
 80      "verdict": "no significant win",
 81      "system_worked": false
 82    },
 83    "mechanism_signature": {
 84      "structural_match": "dynamics controlled pendulum rollout; shared rnn_small",
 85      "idea_grid": [
 86        0.001,
 87        0.003,
 88        0.01
 89      ],
 90      "baseline_grid_union": [
 91        0.001,
 92        0.003,
 93        0.01
 94      ],
 95      "mechanism_signature": {
 96        "predicted": "higher propagated Bellman uncertainty tracks larger NN residuals",
 97        "predicted_sign": "positive correlation",
 98        "observed_mean_corr": 0.0726964430386442,
 99        "observed_corr_per_seed": [
100          0.07096080660950917,
101          -0.07323832720421877,
102          0.18244864224380744,
103          0.13663901810339665,
104          0.0744279022054703,
105          -0.01757320069773833,
106          0.1342003985639885,
107          0.07370630448493865
108        ],
109        "confirmed": false
110      },
111      "n_train": 400,
112      "n_test": 200,
113      "epochs": 15
114    }
115  }
116}