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": 0.4970664903521538
 17      },
 18      {
 19        "cfg": {
 20          "lr": 0.003
 21        },
 22        "mean": 0.06318101449869573
 23      },
 24      {
 25        "cfg": {
 26          "lr": 0.01
 27        },
 28        "mean": 0.011876139731612056
 29      }
 30    ],
 31    "full": {
 32      "mean": 0.011876139731612056,
 33      "std": 0.00487020680273246,
 34      "per_seed": [
 35        0.010697474703192711,
 36        0.011668417602777481,
 37        0.012986297719180584,
 38        0.019935088232159615,
 39        0.005051083862781525,
 40        0.009833427146077156,
 41        0.018368331715464592,
 42        0.006468996871262789
 43      ],
 44      "n": 8
 45    }
 46  },
 47  "idea": {
 48    "mean": 0.0054059389512985945,
 49    "std": 0.001747436411673786,
 50    "per_seed": [
 51      0.007179650943726301,
 52      0.003484317334368825,
 53      0.006277043838053942,
 54      0.004052966833114624,
 55      0.004533923696726561,
 56      0.0029652707744389772,
 57      0.007882867939770222,
 58      0.006871470250189304
 59    ],
 60    "n": 8,
 61    "best_cfg": {
 62      "lr": 0.01
 63    }
 64  },
 65  "comparison": {
 66    "delta_mean": -0.006470200780313462,
 67    "idea_wins": 7,
 68    "n_pairs": 8,
 69    "per_seed_diffs": [
 70      -0.0035178237594664097,
 71      -0.008184100268408656,
 72      -0.006709253881126642,
 73      -0.01588212139904499,
 74      -0.0005171601660549641,
 75      -0.006868156371638179,
 76      -0.01048546377569437,
 77      0.0004024733789265156
 78    ],
 79    "p_value": 0.0156,
 80    "mde": 0.004461024219396654,
 81    "mde_rel_pct": 37.56291455145348,
 82    "verdict": "idea better (significant)",
 83    "system_worked": true
 84  },
 85  "baseline_grid": [
 86    {
 87      "lr": 0.001
 88    },
 89    {
 90      "lr": 0.003
 91    },
 92    {
 93      "lr": 0.01
 94    }
 95  ],
 96  "idea_grid": [
 97    {
 98      "cfg": {
 99        "lr": 0.001
100      },
101      "result": {
102        "mean": 0.1896328409202397,
103        "std": 0.10288670070387347,
104        "per_seed": [
105          0.3404313623905182,
106          0.22487185895442963,
107          0.07244019210338593,
108          0.19320537149906158,
109          0.2565637230873108,
110          0.05929507687687874,
111          0.07232630252838135,
112          0.2979288399219513
113        ],
114        "n": 8
115      }
116    },
117    {
118      "cfg": {
119        "lr": 0.003
120      },
121      "result": {
122        "mean": 0.02868673810735345,
123        "std": 0.012694117890344024,
124        "per_seed": [
125          0.04536343365907669,
126          0.028771521523594856,
127          0.02157784253358841,
128          0.036481499671936035,
129          0.03701217100024223,
130          0.00805759523063898,
131          0.012121920473873615,
132          0.04010792076587677
133        ],
134        "n": 8
135      }
136    },
137    {
138      "cfg": {
139        "lr": 0.01
140      },
141      "result": {
142        "mean": 0.0054059389512985945,
143        "std": 0.001747436411673786,
144        "per_seed": [
145          0.007179650943726301,
146          0.003484317334368825,
147          0.006277043838053942,
148          0.004052966833114624,
149          0.004533923696726561,
150          0.0029652707744389772,
151          0.007882867939770222,
152          0.006871470250189304
153        ],
154        "n": 8,
155        "best_cfg": {
156          "lr": 0.01
157        }
158      }
159    }
160  ],
161  "math_check_exponential_residual": 1.9421449906076305e-06,
162  "protocol_note": "dynamics selected because the idea claims stability/control through norm-preserving recurrent transformations",
163  "mechanism_signature": {
164    "prediction": "KLU recurrent map preserves the norm of each incoming hidden state while dense recurrence need not",
165    "observed_idea_recurrent_norm": 2.192119225859642,
166    "observed_baseline_recurrent_output_norm": 1.5108804553747177,
167    "observed_idea_unitarity_residual": 1.8216120736269659e-06,
168    "confirmed": true
169  }
170}