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": 0.005,
 10      "rho_cap": null
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "rho_cap": null
 17        },
 18        "mean": 0.0012295851483941078
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "rho_cap": null
 24        },
 25        "mean": 0.0006583010253962129
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.005,
 30          "rho_cap": null
 31        },
 32        "mean": 0.0004532121238298714
 33      }
 34    ],
 35    "full": {
 36      "mean": 0.00038185320590855554,
 37      "std": 0.00016136478368296086,
 38      "per_seed": [
 39        0.00046871721860952675,
 40        0.0002616161364130676,
 41        0.00031301137642003596,
 42        0.0007695037638768554,
 43        0.0003513892588671297,
 44        0.0002713301219046116,
 45        0.0002475073270034045,
 46        0.00037175044417381287
 47      ],
 48      "n": 8
 49    }
 50  },
 51  "idea": {
 52    "mean": 0.00016176951703528175,
 53    "std": 2.54928932189329e-05,
 54    "per_seed": [
 55      0.00013927802501711994,
 56      0.0001758444996085018,
 57      0.0001558296353323385,
 58      0.00020030193263664842,
 59      0.0001420884218532592,
 60      0.00012167391105322167,
 61      0.00016640945977997035,
 62      0.00019273025100119412
 63    ],
 64    "n": 8
 65  },
 66  "comparison": {
 67    "delta_mean": -0.00022008368887327379,
 68    "idea_wins": 8,
 69    "n_pairs": 8,
 70    "per_seed_diffs": [
 71      -0.0003294391935924068,
 72      -8.577163680456579e-05,
 73      -0.00015718174108769745,
 74      -0.000569201831240207,
 75      -0.00020930083701387048,
 76      -0.00014965621085138991,
 77      -8.109786722343415e-05,
 78      -0.00017902019317261875
 79    ],
 80    "p_value": 0.0081,
 81    "mde": 0.0001347953881261106,
 82    "mde_rel_pct": 35.3003159435018,
 83    "verdict": "idea better (significant)",
 84    "system_worked": true
 85  },
 86  "mechanism_signature": {
 87    "signature": {
 88      "prediction": "projected recurrent operator-gain proxy stays <= 0.90",
 89      "predicted_max": 0.7689717411994934,
 90      "observed": [
 91        {
 92          "seed": 0,
 93          "predicted_gain_proxy": 0.7254310846328735,
 94          "observed_output_gain": 0.0707114040851593
 95        },
 96        {
 97          "seed": 1,
 98          "predicted_gain_proxy": 0.7678388953208923,
 99          "observed_output_gain": 0.09257301688194275
100        },
101        {
102          "seed": 2,
103          "predicted_gain_proxy": 0.7689717411994934,
104          "observed_output_gain": 0.07351840287446976
105        },
106        {
107          "seed": 3,
108          "predicted_gain_proxy": 0.7672179341316223,
109          "observed_output_gain": 0.07399722933769226
110        }
111      ],
112      "confirmed": true
113    },
114    "idea_sweep": {
115      "0.001": {
116        "mean": 0.00032572537929809187,
117        "std": 6.515504960120716e-05,
118        "per_seed": [
119          0.00024285227118525654,
120          0.00030959199648350477,
121          0.0003330624313093722,
122          0.00036215895670466125,
123          0.00030998658621683717,
124          0.0002514962980058044,
125          0.0003306802245788276,
126          0.00046597426990047097
127        ],
128        "n": 8
129      },
130      "0.003": {
131        "mean": 0.0002240399771835655,
132        "std": 3.985152266293653e-05,
133        "per_seed": [
134          0.00017446346464566886,
135          0.00022571365116164088,
136          0.00022476879530586302,
137          0.0002838608343154192,
138          0.00020102885900996625,
139          0.00017465419659856707,
140          0.00022299615375231951,
141          0.0002848338626790792
142        ],
143        "n": 8
144      },
145      "0.005": {
146        "mean": 0.00016176951703528175,
147        "std": 2.54928932189329e-05,
148        "per_seed": [
149          0.00013927802501711994,
150          0.0001758444996085018,
151          0.0001558296353323385,
152          0.00020030193263664842,
153          0.0001420884218532592,
154          0.00012167391105322167,
155          0.00016640945977997035,
156          0.00019273025100119412
157        ],
158        "n": 8
159      }
160    },
161    "selected_idea_lr": 0.005,
162    "structure": "GRU recurrent dynamics; spectral-gain projection is the only intervention"
163  },
164  "baseline_sweep_union_note": "Baseline evaluated all idea learning rates; baseline method has no additional decisive knob."
165}