Monotone Compositional Reachability Critic / 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      "epochs": 10
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "epochs": 10
 17        },
 18        "mean": 0.018210270907729864
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "epochs": 10
 24        },
 25        "mean": 0.009255725424736738
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.006,
 30          "epochs": 10
 31        },
 32        "mean": 0.004965097992680967
 33      }
 34    ],
 35    "full": {
 36      "mean": 0.005481509841047227,
 37      "std": 0.0017443505414011541,
 38      "per_seed": [
 39        0.0044007012620568275,
 40        0.006472812034189701,
 41        0.004313930869102478,
 42        0.004672947805374861,
 43        0.003332370426505804,
 44        0.008815406821668148,
 45        0.007345432415604591,
 46        0.004498477093875408
 47      ],
 48      "n": 8
 49    }
 50  },
 51  "idea": {
 52    "mean": 0.0049165511154569685,
 53    "std": 0.0019956794930531424,
 54    "per_seed": [
 55      0.005324999801814556,
 56      0.0023787340614944696,
 57      0.002438431140035391,
 58      0.00647375825792551,
 59      0.0034786139149218798,
 60      0.0063143507577478886,
 61      0.0044753411784768105,
 62      0.008448179811239243
 63    ],
 64    "n": 8,
 65    "sweep": [
 66      {
 67        "cfg": {
 68          "lr": 0.001,
 69          "epochs": 10
 70        },
 71        "mean": 0.03285314375534654
 72      },
 73      {
 74        "cfg": {
 75          "lr": 0.003,
 76          "epochs": 10
 77        },
 78        "mean": 0.0049165511154569685
 79      },
 80      {
 81        "cfg": {
 82          "lr": 0.006,
 83          "epochs": 10
 84        },
 85        "mean": 0.0050386641814839095
 86      }
 87    ]
 88  },
 89  "comparison": {
 90    "delta_mean": -0.0005649587255902588,
 91    "idea_wins": 4,
 92    "n_pairs": 8,
 93    "per_seed_diffs": [
 94      0.0009242985397577286,
 95      -0.0040940779726952314,
 96      -0.0018754997290670872,
 97      0.0018008104525506496,
 98      0.0001462434884160757,
 99      -0.0025010560639202595,
100      -0.002870091237127781,
101      0.003949702717363834
102    ],
103    "p_value": 0.56675,
104    "mde": 0.0022775908803314825,
105    "mde_rel_pct": 41.55042946883326,
106    "verdict": "no significant win",
107    "system_worked": false
108  },
109  "mechanism_signature": {
110    "prediction": "monotone aggregator has nonnegative derivatives; increasing a primitive cannot lower composite value",
111    "baseline_observed": {
112      "samples": 128,
113      "derivative_nonnegative_fraction": 0.0,
114      "mean_derivatives": [
115        -0.4277612566947937,
116        -0.4763301610946655
117      ],
118      "positive_primitive_perturbation_mean_response": -0.04277614504098892
119    },
120    "idea_observed": {
121      "samples": 128,
122      "derivative_nonnegative_fraction": 1.0,
123      "mean_derivatives": [
124        0.5155500173568726,
125        0.5073314309120178
126      ],
127      "positive_primitive_perturbation_mean_response": 0.051555000245571136
128    },
129    "confirmed": true,
130    "measurement_note": "All quantities were computed from trained benchmark models on held-out dynamics inputs."
131  },
132  "custom_track": null
133}