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": 0.0019389573426451534
 17      },
 18      {
 19        "cfg": {
 20          "lr": 0.003
 21        },
 22        "mean": 0.000923123603570275
 23      },
 24      {
 25        "cfg": {
 26          "lr": 0.006
 27        },
 28        "mean": 0.0006749839230906218
 29      }
 30    ],
 31    "full": {
 32      "mean": 0.0006854402672615834,
 33      "std": 0.00019420169122248883,
 34      "per_seed": [
 35        0.0006784203578718007,
 36        0.0006286462303251028,
 37        0.0005217542639002204,
 38        0.0008711148402653635,
 39        0.0005711782141588628,
 40        0.000547860749065876,
 41        0.0005476096412166953,
 42        0.0011169378412887454
 43      ],
 44      "n": 8
 45    }
 46  },
 47  "idea": {
 48    "mean": 0.0006295234306890052,
 49    "std": 0.0001478583427936034,
 50    "per_seed": [
 51      0.0006697725621052086,
 52      0.0006321860128082335,
 53      0.000446179648861289,
 54      0.0007750660297460854,
 55      0.0005767291877418756,
 56      0.00047575801727361977,
 57      0.0005408408469520509,
 58      0.0009196551400236785
 59    ],
 60    "n": 8
 61  },
 62  "comparison": {
 63    "delta_mean": -5.591683657257818e-05,
 64    "idea_wins": 6,
 65    "n_pairs": 8,
 66    "per_seed_diffs": [
 67      -8.647795766592026e-06,
 68      3.5397824831306934e-06,
 69      -7.557461503893137e-05,
 70      -9.604881051927805e-05,
 71      5.550973583012819e-06,
 72      -7.210273179225624e-05,
 73      -6.768794264644384e-06,
 74      -0.00019728270126506686
 75    ],
 76    "p_value": 0.0443,
 77    "mde": 5.845099158159299e-05,
 78    "mde_rel_pct": 8.527510619577656,
 79    "verdict": "idea better (significant)",
 80    "system_worked": true
 81  },
 82  "mechanism_signature": {
 83    "math_sanity": {
 84      "predicted_damping_shift": 1.25,
 85      "observed_shift": 1.2500000000000004,
 86      "abs_error": 4.440892098500626e-16,
 87      "passed": true
 88    },
 89    "trained_jacobian": {
 90      "quantity": "trained-model local input Jacobian norm",
 91      "predicted": "periodic modulation reduces perturbation growth",
 92      "baseline_mean": 0.4034101516008377,
 93      "idea_mean": 0.40224631130695343,
 94      "relative_change": -0.0028850049738853934,
 95      "confirmed": true
 96    },
 97    "idea_cfg": {
 98      "lr": 0.006
 99    },
100    "idea_sweep": [
101      {
102        "cfg": {
103          "lr": 0.001
104        },
105        "mean": 0.0019525550596881658
106      },
107      {
108        "cfg": {
109          "lr": 0.003
110        },
111        "mean": 0.000914391697733663
112      },
113      {
114        "cfg": {
115          "lr": 0.006
116        },
117        "mean": 0.0006308010633802041
118      }
119    ]
120  },
121  "selection_note": "Baseline and idea share rnn_small, data, epochs, batch, AdamW weight decay, and union learning-rate grid; only periodic LR modulation differs."
122}