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    },
 11    "sweep": [
 12      {
 13        "cfg": {
 14          "lr": 0.001
 15        },
 16        "mean": 0.0012489092478062958
 17      },
 18      {
 19        "cfg": {
 20          "lr": 0.003
 21        },
 22        "mean": 0.0007282123697223142
 23      },
 24      {
 25        "cfg": {
 26          "lr": 0.01
 27        },
 28        "mean": 0.0002650022797752172
 29      }
 30    ],
 31    "full": {
 32      "mean": 0.00020160043823125307,
 33      "std": 0.00011406837376812385,
 34      "per_seed": [
 35        0.0002867693838197738,
 36        0.00022152646852191538,
 37        0.00010455308074597269,
 38        0.00044716018601320684,
 39        8.912176417652518e-05,
 40        0.0002234217245131731,
 41        0.00013277865946292877,
 42        0.00010747223859652877
 43      ],
 44      "n": 8
 45    }
 46  },
 47  "idea": {
 48    "mean": 0.0008653015684103593,
 49    "std": 0.00041340883077491884,
 50    "per_seed": [
 51      0.0006581585621461272,
 52      0.0009649579296819866,
 53      0.0017265599453821778,
 54      0.0007086805999279022,
 55      0.000453480111900717,
 56      0.001295149908401072,
 57      0.0006053093820810318,
 58      0.0005101161077618599
 59    ],
 60    "n": 8
 61  },
 62  "comparison": {
 63    "delta_mean": 0.0006637011301791063,
 64    "idea_wins": 0,
 65    "n_pairs": 8,
 66    "per_seed_diffs": [
 67      0.00037138917832635343,
 68      0.0007434314611600712,
 69      0.0016220068646362051,
 70      0.0002615204139146954,
 71      0.00036435834772419184,
 72      0.001071728183887899,
 73      0.00047253072261810303,
 74      0.0004026438691653311
 75    ],
 76    "p_value": 0.0081,
 77    "mde": 0.0003912591628540673,
 78    "mde_rel_pct": 194.07654382440347,
 79    "verdict": "idea worse (significant)",
 80    "system_worked": false
 81  },
 82  "mechanism_signature": {
 83    "quantity": "finite perturbation amplification of trained memory features",
 84    "predicted_alpha_bound": 1.0,
 85    "observed_amplification": 5.219648766797036,
 86    "n_models": 8,
 87    "tolerance": "observed <= alpha + 0.15",
 88    "confirmed": false
 89  },
 90  "idea_sweep": [
 91    {
 92      "cfg": {
 93        "lr": 0.01,
 94        "alpha": 0.8
 95      },
 96      "mean": 0.0010894885854213499
 97    },
 98    {
 99      "cfg": {
100        "lr": 0.01,
101        "alpha": 0.9
102      },
103      "mean": 0.0009876549775071908
104    },
105    {
106      "cfg": {
107        "lr": 0.01,
108        "alpha": 1.0
109      },
110      "mean": 0.0008653015684103593
111    }
112  ],
113  "protocol_notes": {
114    "epochs": 15,
115    "n_train": 800,
116    "n_test": 200,
117    "matched_task": "controlled damped pendulum horizon-8",
118    "architecture_difference": "GRU recurrence versus two-column A/B memory"
119  }
120}