1{
  2  "bench_version": 1,
  3  "track": "dynamics",
  4  "model": "switching_rnn_small",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.01,
 10      "weight_decay": 0.0
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "weight_decay": 0.0
 17        },
 18        "mean": 0.6787904053926468
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "weight_decay": 0.0
 24        },
 25        "mean": 0.33453836664557457
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.01,
 30          "weight_decay": 0.0
 31        },
 32        "mean": 0.06969651952385902
 33      }
 34    ],
 35    "full": {
 36      "mean": 0.060021675657480955,
 37      "std": 0.015156924357368727,
 38      "per_seed": [
 39        0.08692938834428787,
 40        0.06139367073774338,
 41        0.07089249789714813,
 42        0.059570521116256714,
 43        0.048601675778627396,
 44        0.06793582439422607,
 45        0.051944129168987274,
 46        0.0329056978225708
 47      ],
 48      "n": 8
 49    }
 50  },
 51  "idea": {
 52    "mean": 0.06696924939751625,
 53    "std": 0.012785122715736058,
 54    "per_seed": [
 55      0.0705910176038742,
 56      0.057215094566345215,
 57      0.08217969536781311,
 58      0.06861463189125061,
 59      0.06553272157907486,
 60      0.0647139772772789,
 61      0.0850575640797615,
 62      0.0418492928147316
 63    ],
 64    "n": 8
 65  },
 66  "comparison": {
 67    "delta_mean": 0.0069475737400352955,
 68    "idea_wins": 3,
 69    "n_pairs": 8,
 70    "per_seed_diffs": [
 71      -0.016338370740413666,
 72      -0.004178576171398163,
 73      0.011287197470664978,
 74      0.009044110774993896,
 75      0.016931045800447464,
 76      -0.003221847116947174,
 77      0.03311343491077423,
 78      0.008943594992160797
 79    ],
 80    "p_value": 0.23195,
 81    "mde": 0.012566084314756818,
 82    "mde_rel_pct": 20.935910530832725,
 83    "verdict": "no significant win",
 84    "system_worked": false
 85  },
 86  "mechanism_signature": {
 87    "signature": {
 88      "certificate": "V(z)=||z||^2; one-node path-complete graph; four all-mode self-edges",
 89      "prediction": "edge contraction ratio <= rho",
 90      "rho": 0.81,
 91      "observed_max_zero_input_ratio": 0.5740312337875366,
 92      "observed_mean_ratio": 0.19770055771505213,
 93      "confirmed": true,
 94      "n_trained_models": 8
 95    },
 96    "idea_cfg": {
 97      "lr": 0.01,
 98      "lambda_contract": 0.2
 99    },
100    "training_budget": "400 train/400 test samples, 5 epochs, batch 128; equal on both systems"
101  },
102  "idea_sweep": [
103    {
104      "cfg": {
105        "lr": 0.01,
106        "lambda_contract": 0.01
107      },
108      "result": {
109        "mean": 0.06696925032883883,
110        "std": 0.012785122979561253,
111        "per_seed": [
112          0.0705910250544548,
113          0.057215094566345215,
114          0.08217969536781311,
115          0.06861463189125061,
116          0.06553272157907486,
117          0.0647139772772789,
118          0.0850575640797615,
119          0.0418492928147316
120        ],
121        "n": 8
122      }
123    },
124    {
125      "cfg": {
126        "lr": 0.01,
127        "lambda_contract": 0.05
128      },
129      "result": {
130        "mean": 0.06696925032883883,
131        "std": 0.012785122979561253,
132        "per_seed": [
133          0.0705910250544548,
134          0.057215094566345215,
135          0.08217969536781311,
136          0.06861463189125061,
137          0.06553272157907486,
138          0.0647139772772789,
139          0.0850575640797615,
140          0.0418492928147316
141        ],
142        "n": 8
143      }
144    },
145    {
146      "cfg": {
147        "lr": 0.01,
148        "lambda_contract": 0.2
149      },
150      "result": {
151        "mean": 0.06696924939751625,
152        "std": 0.012785122715736058,
153        "per_seed": [
154          0.0705910176038742,
155          0.057215094566345215,
156          0.08217969536781311,
157          0.06861463189125061,
158          0.06553272157907486,
159          0.0647139772772789,
160          0.0850575640797615,
161          0.0418492928147316
162        ],
163        "n": 8
164      }
165    }
166  ],
167  "custom_track": null,
168  "device": "cuda"
169}