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": 0.0007738938584225252
 17      },
 18      {
 19        "cfg": {
 20          "lr": 0.003
 21        },
 22        "mean": 0.0004595852587954141
 23      },
 24      {
 25        "cfg": {
 26          "lr": 0.006
 27        },
 28        "mean": 0.00018638806068338454
 29      }
 30    ],
 31    "full": {
 32      "mean": 0.00014496273979602847,
 33      "std": 5.6302020060222334e-05,
 34      "per_seed": [
 35        0.00022238190285861492,
 36        0.0002155364491045475,
 37        0.00011000921949744225,
 38        0.00019762467127293348,
 39        0.00014344952069222927,
 40        8.176809205906466e-05,
 41        6.902739551151171e-05,
 42        0.00011990466737188399
 43      ],
 44      "n": 8
 45    }
 46  },
 47  "idea": {
 48    "mean": 0.00014406874470296316,
 49    "std": 5.6322857127222856e-05,
 50    "per_seed": [
 51      0.00022284468286670744,
 52      0.00021447740436997265,
 53      0.00010853678395505995,
 54      0.00019537980551831424,
 55      0.00014281165204010904,
 56      8.075518417172134e-05,
 57      6.867131014587358e-05,
 58      0.00011907313455594704
 59    ],
 60    "n": 8
 61  },
 62  "comparison": {
 63    "delta_mean": -8.93995093065314e-07,
 64    "idea_wins": 7,
 65    "n_pairs": 8,
 66    "per_seed_diffs": [
 67      4.627800080925226e-07,
 68      -1.0590447345748544e-06,
 69      -1.4724355423823e-06,
 70      -2.2448657546192408e-06,
 71      -6.378686521202326e-07,
 72      -1.0129078873433173e-06,
 73      -3.5608536563813686e-07,
 74      -8.315328159369528e-07
 75    ],
 76    "p_value": 0.0224,
 77    "mde": 6.629623320875621e-07,
 78    "mde_rel_pct": 0.4573329208735921,
 79    "verdict": "idea better (significant)",
 80    "system_worked": true
 81  },
 82  "mechanism_signature": {
 83    "idea_lr_sweep": [
 84      {
 85        "cfg": {
 86          "lr": 0.0015
 87        },
 88        "mean": 0.0007737674750387669
 89      },
 90      {
 91        "cfg": {
 92          "lr": 0.003
 93        },
 94        "mean": 0.0004588514202623628
 95      },
 96      {
 97        "cfg": {
 98          "lr": 0.006
 99        },
100        "mean": 0.00018530966917751357
101      }
102    ],
103    "chosen_idea_cfg": {
104      "lr": 0.006
105    },
106    "predicted_max_jacobian": 0.0,
107    "observed_mean_abs_alpha_over_tau": 0.049265923735219985,
108    "observed_residual_disturbance_corr_abs": 0.999999999999998,
109    "n_trained_models": 8,
110    "confirmed": true,
111    "interpretation": "The matched bench has no latent actuator error; the simplicity prior should identify zero disturbance."
112  }
113}