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      "weight_decay": 0.0001
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "weight_decay": 0.0
 17        },
 18        "mean": 0.0061489523795899
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.001,
 23          "weight_decay": 0.0001
 24        },
 25        "mean": 0.006141080608358607
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.003,
 30          "weight_decay": 0.0
 31        },
 32        "mean": 0.0013683813158422709
 33      },
 34      {
 35        "cfg": {
 36          "lr": 0.003,
 37          "weight_decay": 0.0001
 38        },
 39        "mean": 0.0013369002845138311
 40      },
 41      {
 42        "cfg": {
 43          "lr": 0.01,
 44          "weight_decay": 0.0
 45        },
 46        "mean": 0.0007866571540944278
 47      },
 48      {
 49        "cfg": {
 50          "lr": 0.01,
 51          "weight_decay": 0.0001
 52        },
 53        "mean": 0.0007708658231422305
 54      }
 55    ],
 56    "full": {
 57      "mean": 0.0006831642931501847,
 58      "std": 0.00017069130263284838,
 59      "per_seed": [
 60        0.000798072200268507,
 61        0.0005629754159599543,
 62        0.0006793037755414844,
 63        0.0010431119007989764,
 64        0.0005458718515001237,
 65        0.0004460137861315161,
 66        0.000679348420817405,
 67        0.0007106169941835105
 68      ],
 69      "n": 8
 70    }
 71  },
 72  "idea": {
 73    "mean": 0.0006341310436255299,
 74    "std": 0.00015214794700785093,
 75    "per_seed": [
 76      0.0008724817307665944,
 77      0.00042524098535068333,
 78      0.0005125870229676366,
 79      0.0008023462723940611,
 80      0.0004432320420164615,
 81      0.000654117320664227,
 82      0.0006684261607006192,
 83      0.0006946168141439557
 84    ],
 85    "n": 8,
 86    "sweep": [
 87      {
 88        "cfg": {
 89          "lr": 0.01,
 90          "weight_decay": 0.0001,
 91          "alpha": 0.1
 92        },
 93        "mean": 0.0008600900182500482
 94      },
 95      {
 96        "cfg": {
 97          "lr": 0.01,
 98          "weight_decay": 0.0001,
 99          "alpha": 0.3
100        },
101        "mean": 0.0006531640028697439
102      },
103      {
104        "cfg": {
105          "lr": 0.01,
106          "weight_decay": 0.0001,
107          "alpha": 1.0
108        },
109        "mean": 0.0008107436296995729
110      },
111      {
112        "cfg": {
113          "lr": 0.001,
114          "weight_decay": 0.0001,
115          "alpha": 0.3
116        },
117        "mean": 0.006068007613066584
118      },
119      {
120        "cfg": {
121          "lr": 0.003,
122          "weight_decay": 0.0001,
123          "alpha": 0.3
124        },
125        "mean": 0.0012003584706690162
126      }
127    ]
128  },
129  "comparison": {
130    "delta_mean": -4.903324952465482e-05,
131    "idea_wins": 6,
132    "n_pairs": 8,
133    "per_seed_diffs": [
134      7.44095304980874e-05,
135      -0.00013773443060927093,
136      -0.00016671675257384777,
137      -0.00024076562840491533,
138      -0.00010263980948366225,
139      0.0002081035345327109,
140      -1.0922260116785765e-05,
141      -1.6000180039554834e-05
142    ],
143    "p_value": 0.36915,
144    "mde": 0.00012044817185346833,
145    "mde_rel_pct": 17.630923199756488,
146    "verdict": "no significant win",
147    "system_worked": false
148  },
149  "mechanism_signature": {
150    "prediction": "dual Hankel training yields finite future-block reconstruction and latent transition residual",
151    "observed_future_mse": 0.27331745624542236,
152    "observed_past_mse": 0.21698212623596191,
153    "observed_transition_mse": 0.01214107871055603,
154    "latent_spectral_radius": 0.9021337628364563,
155    "confirmed": true
156  }
157}