PAC transition-cover training monitor / 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": 0.006
 10    },
 11    "sweep": [
 12      {
 13        "cfg": {
 14          "lr": 0.001
 15        },
 16        "mean": 0.46847324818372726
 17      },
 18      {
 19        "cfg": {
 20          "lr": 0.003
 21        },
 22        "mean": 0.028258539736270905
 23      },
 24      {
 25        "cfg": {
 26          "lr": 0.006
 27        },
 28        "mean": 0.021621495485305786
 29      }
 30    ],
 31    "full": {
 32      "mean": 0.02255184983368963,
 33      "std": 0.009104915744183411,
 34      "per_seed": [
 35        0.03419062867760658,
 36        0.01129467599093914,
 37        0.018067732453346252,
 38        0.02293294481933117,
 39        0.03928421065211296,
 40        0.0164501890540123,
 41        0.023551812395453453,
 42        0.014642604626715183
 43      ],
 44      "n": 8
 45    }
 46  },
 47  "idea": {
 48    "mean": 0.3601643927395344,
 49    "std": 0.08035137726894388,
 50    "per_seed": [
 51      0.27401673793792725,
 52      0.4190290868282318,
 53      0.334565132856369,
 54      0.48182761669158936,
 55      0.21452373266220093,
 56      0.3759009838104248,
 57      0.3592603802680969,
 58      0.42219147086143494
 59    ],
 60    "n": 8
 61  },
 62  "comparison": {
 63    "delta_mean": 0.33761254290584475,
 64    "idea_wins": 0,
 65    "n_pairs": 8,
 66    "per_seed_diffs": [
 67      0.23982610926032066,
 68      0.40773441083729267,
 69      0.31649740040302277,
 70      0.4588946718722582,
 71      0.17523952201008797,
 72      0.3594507947564125,
 73      0.33570856787264347,
 74      0.40754886623471975
 75    ],
 76    "p_value": 0.0081,
 77    "mde": 0.07829229882889692,
 78    "mde_rel_pct": 347.1657509528911,
 79    "verdict": "idea worse (significant)",
 80    "system_worked": false
 81  },
 82  "mechanism_signature": {
 83    "pac": {
 84      "epsilon": 0.1,
 85      "delta": 0.1
 86    },
 87    "best_idea_lr": 0.006,
 88    "signature": {
 89      "epsilon": 0.1,
 90      "delta": 0.1,
 91      "required_n": 24,
 92      "pac_counts": [
 93        31,
 94        95,
 95        43,
 96        33,
 97        47,
 98        48,
 99        80,
100        7
101      ],
102      "envelope_radius": 1.101591670513153,
103      "heldout_miss_rate_mean": 0.30166666666666664,
104      "heldout_miss_rate_by_cell": [
105        1.0,
106        0.08,
107        0.0,
108        0.0,
109        0.0,
110        0.0,
111        0.3333333333333333,
112        1.0
113      ],
114      "prediction": "PAC-ready cells should have held-out miss rate <= epsilon",
115      "confirmed": false,
116      "baseline_test_mse": 0.06002972647547722,
117      "idea_test_mse": 0.45216846466064453
118    }
119  },
120  "idea_sweep": [
121    {
122      "cfg": {
123        "lr": 0.001
124      },
125      "mean": 0.6608457192778587,
126      "std": 0.0759199143760643,
127      "n": 8
128    },
129    {
130      "cfg": {
131        "lr": 0.003
132      },
133      "mean": 0.5322897769510746,
134      "std": 0.07522369454464452,
135      "n": 8
136    },
137    {
138      "cfg": {
139        "lr": 0.006
140      },
141      "mean": 0.3601643927395344,
142      "std": 0.08035137726894388,
143      "n": 8
144    }
145  ],
146  "protocol": {
147    "paired_seeds": [
148      0,
149      1,
150      2,
151      3,
152      4,
153      5,
154      6,
155      7
156    ],
157    "epochs": 3,
158    "n_train": 400,
159    "structural_match": "dynamics transition model / stability-control"
160  }
161}