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": 0.0032877353951334953
 17      },
 18      {
 19        "cfg": {
 20          "lr": 0.003
 21        },
 22        "mean": 0.0012035724357701838
 23      },
 24      {
 25        "cfg": {
 26          "lr": 0.01
 27        },
 28        "mean": 0.0006944253206408272
 29      }
 30    ],
 31    "full": {
 32      "mean": 0.0006787698075640947,
 33      "std": 0.00018311611365834393,
 34      "per_seed": [
 35        0.0008624937618151307,
 36        0.0008604836766608059,
 37        0.000360298523446545,
 38        0.0006422948790714145,
 39        0.0006247925339266658,
 40        0.0009110610117204487,
 41        0.0004747614439111203,
 42        0.0006939726299606264
 43      ],
 44      "n": 8
 45    }
 46  },
 47  "idea": {
 48    "mean": 0.0002450490219416679,
 49    "std": 0.0001979632966612432,
 50    "per_seed": [
 51      0.00021305571135599166,
 52      0.0002364848623983562,
 53      0.0007525552064180374,
 54      0.0001690309727564454,
 55      0.00017366997781209648,
 56      8.526047167833894e-05,
 57      0.0002184970217058435,
 58      0.0001118379514082335
 59    ],
 60    "n": 8
 61  },
 62  "comparison": {
 63    "delta_mean": -0.0004337207856224268,
 64    "idea_wins": 7,
 65    "n_pairs": 8,
 66    "per_seed_diffs": [
 67      -0.000649438050459139,
 68      -0.0006239988142624497,
 69      0.0003922566829714924,
 70      -0.00047326390631496906,
 71      -0.0004511225561145693,
 72      -0.0008258005400421098,
 73      -0.0002562644222052768,
 74      -0.0005821346785523929
 75    ],
 76    "p_value": 0.0231,
 77    "mde": 0.00031191512145065987,
 78    "mde_rel_pct": 45.953004682107405,
 79    "verdict": "idea better (significant)",
 80    "system_worked": true
 81  },
 82  "math_check": {
 83    "rls_batch_max_abs_error": 5.551115123125783e-17,
 84    "predicted_adaptation_timescale": 14.285714285714295
 85  },
 86  "idea_sweep": [
 87    {
 88      "cfg": {
 89        "lr": 0.001
 90      },
 91      "result": {
 92        "mean": 0.00044900484863319434,
 93        "std": 0.00010002042742588506,
 94        "per_seed": [
 95          0.0007017979514785111,
 96          0.0004040250787511468,
 97          0.00040901819011196494,
 98          0.00039653119165450335,
 99          0.00039476461824961007,
100          0.0004773136170115322,
101          0.00043600559001788497,
102          0.00037258255179040134
103        ],
104        "n": 8
105      }
106    },
107    {
108      "cfg": {
109        "lr": 0.003
110      },
111      "result": {
112        "mean": 0.00025152570015052333,
113        "std": 5.887341333514287e-05,
114        "per_seed": [
115          0.0002252471458632499,
116          0.00016258451796602458,
117          0.0003189823473803699,
118          0.0002701112534850836,
119          0.00025544039090164006,
120          0.00024129776284098625,
121          0.00035137461964040995,
122          0.00018716756312642246
123        ],
124        "n": 8
125      }
126    },
127    {
128      "cfg": {
129        "lr": 0.01
130      },
131      "result": {
132        "mean": 0.0002450490219416679,
133        "std": 0.0001979632966612432,
134        "per_seed": [
135          0.00021305571135599166,
136          0.0002364848623983562,
137          0.0007525552064180374,
138          0.0001690309727564454,
139          0.00017366997781209648,
140          8.526047167833894e-05,
141          0.0002184970217058435,
142          0.0001118379514082335
143        ],
144        "n": 8
145      }
146    }
147  ],
148  "protocol_note": "Baseline sweep uses same lr union; final comparison is eight paired seeds.",
149  "mechanism_signature": {
150    "rho_trained_A": 0.7977778911590576,
151    "predicted_asymptotic_norm_ratio": 0.7977778911590576,
152    "observed_mean_last3_norm_ratio": 0.8155918888124241,
153    "horizon_norms": [
154      4.329426288604736,
155      1.5744551420211792,
156      0.8551374077796936,
157      0.697711169719696,
158      0.6099732518196106,
159      0.518440306186676,
160      0.42884668707847595,
161      0.3488800525665283,
162      0.2812177240848541
163    ],
164    "confirmed": true,
165    "source": "trained benchmark Koopman model"
166  }
167}