Zero-Crossing Reset Integral Optimizer / bench_report.json

Mechanism confirmed, baseline not beaten

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    },
 11    "sweep": [
 12      {
 13        "cfg": {
 14          "lr": 0.0015
 15        },
 16        "mean": 0.0011411938758101314
 17      },
 18      {
 19        "cfg": {
 20          "lr": 0.003
 21        },
 22        "mean": 0.0008498737297486514
 23      },
 24      {
 25        "cfg": {
 26          "lr": 0.006
 27        },
 28        "mean": 0.0006388214169419371
 29      }
 30    ],
 31    "full": {
 32      "mean": 0.0006619469895667862,
 33      "std": 0.00021167023064506447,
 34      "per_seed": [
 35        0.0005903388955630362,
 36        0.0005036640213802457,
 37        0.0004466140235308558,
 38        0.0010146687272936106,
 39        0.0006089333328418434,
 40        0.00040170771535485983,
 41        0.0008167694904841483,
 42        0.00091287971008569
 43      ],
 44      "n": 8
 45    },
 46    "sweep_union": [
 47      0.0015,
 48      0.003,
 49      0.006
 50    ],
 51    "method": "Adam"
 52  },
 53  "idea": {
 54    "mean": 0.036549699725583196,
 55    "std": 0.01617539040826831,
 56    "per_seed": [
 57      0.018661078065633774,
 58      0.03602646663784981,
 59      0.04254741966724396,
 60      0.07239222526550293,
 61      0.02369782328605652,
 62      0.027904292568564415,
 63      0.025326665490865707,
 64      0.045841626822948456
 65    ],
 66    "n": 8,
 67    "sweep": [
 68      {
 69        "cfg": {
 70          "lr": 0.0015
 71        },
 72        "mean": 0.4790512025356293
 73      },
 74      {
 75        "cfg": {
 76          "lr": 0.003
 77        },
 78        "mean": 0.26442450284957886
 79      },
 80      {
 81        "cfg": {
 82          "lr": 0.006
 83        },
 84        "mean": 0.04240679740905762
 85      }
 86    ],
 87    "best_cfg": {
 88      "lr": 0.006
 89    },
 90    "method": "reset_PI"
 91  },
 92  "comparison": {
 93    "delta_mean": 0.03588775273601641,
 94    "idea_wins": 0,
 95    "n_pairs": 8,
 96    "per_seed_diffs": [
 97      0.018070739170070738,
 98      0.03552280261646956,
 99      0.0421008056437131,
100      0.07137755653820932,
101      0.023088889953214675,
102      0.027502584853209555,
103      0.02450989600038156,
104      0.044928747112862766
105    ],
106    "p_value": 0.0081,
107    "mde": 0.014347581754178641,
108    "mde_rel_pct": 2167.481985765729,
109    "verdict": "idea worse (significant)",
110    "system_worked": false
111  },
112  "mechanism_signature": {
113    "prediction": "sign crossings trigger integral resets and dwell limits reset frequency",
114    "observed_mean_sign_flip_events": 427.875,
115    "observed_mean_resets": 59.25,
116    "observed_reset_fraction_of_flip_events": 0.13847502191060473,
117    "dwell_steps": 8,
118    "confirmed": true
119  },
120  "track_justification": "The idea targets stability/control dynamics; the built-in actuated pendulum rollout is structurally matched."
121}