1{
  2  "bench_version": 1,
  3  "track": "sequence",
  4  "model": "custom_recurrent_state",
  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": 1.100467175245285
 17      },
 18      {
 19        "cfg": {
 20          "lr": 0.003
 21        },
 22        "mean": 1.0823257267475128
 23      },
 24      {
 25        "cfg": {
 26          "lr": 0.01
 27        },
 28        "mean": 1.043792873620987
 29      }
 30    ],
 31    "full": {
 32      "mean": 1.0421477407217026,
 33      "std": 0.10101764442672699,
 34      "per_seed": [
 35        1.1993988752365112,
 36        0.8459368944168091,
 37        1.0848031044006348,
 38        1.0450326204299927,
 39        1.0629944801330566,
 40        1.080702543258667,
 41        1.0881637334823608,
 42        0.9301496744155884
 43      ],
 44      "n": 8
 45    }
 46  },
 47  "idea": {
 48    "mean": 1.0151974856853485,
 49    "std": 0.08790083123237098,
 50    "per_seed": [
 51      1.1760079860687256,
 52      0.8348132967948914,
 53      1.027022123336792,
 54      0.9901270866394043,
 55      1.0220698118209839,
 56      0.9881404042243958,
 57      1.0568915605545044,
 58      1.0265076160430908
 59    ],
 60    "n": 8
 61  },
 62  "comparison": {
 63    "delta_mean": -0.026950255036354065,
 64    "idea_wins": 7,
 65    "n_pairs": 8,
 66    "per_seed_diffs": [
 67      -0.023390889167785645,
 68      -0.011123597621917725,
 69      -0.05778098106384277,
 70      -0.05490553379058838,
 71      -0.040924668312072754,
 72      -0.09256213903427124,
 73      -0.031272172927856445,
 74      0.09635794162750244
 75    ],
 76    "p_value": 0.21995,
 77    "mde": 0.04657542404598468,
 78    "mde_rel_pct": 4.46917670365341,
 79    "verdict": "no significant win",
 80    "system_worked": false
 81  },
 82  "mechanism_signature": {
 83    "prediction": "reversal output difference is explained by trained area-head contribution",
 84    "n": 64,
 85    "observed_output_reversal_rms": 0.28506407141685486,
 86    "predicted_area_head_rms": 0.28506410121917725,
 87    "observed_to_predicted_rms_ratio": 0.9999998954535234,
 88    "correlation": 0.9999999999999685,
 89    "confirmed": true,
 90    "model_test_mse": 1.1760079860687256,
 91    "idea_sweep": {
 92      "best_cfg": {
 93        "lr": 0.01
 94      },
 95      "sweep": [
 96        {
 97          "cfg": {
 98            "lr": 0.001
 99          },
100          "mean": 1.0460671931505203
101        },
102        {
103          "cfg": {
104            "lr": 0.003
105          },
106          "mean": 1.0292403250932693
107        },
108        {
109          "cfg": {
110            "lr": 0.01
111          },
112          "mean": 1.0069926232099533
113        }
114      ],
115      "full": {
116        "mean": 1.0151974856853485,
117        "std": 0.08790083123237098,
118        "per_seed": [
119          1.1760079860687256,
120          0.8348132967948914,
121          1.027022123336792,
122          0.9901270866394043,
123          1.0220698118209839,
124          0.9881404042243958,
125          1.0568915605545044,
126          1.0265076160430908
127        ],
128        "n": 8
129      }
130    }
131  },
132  "protocol_notes": {
133    "architecture": "matched projection/state/readout; only area channels replace additive state",
134    "epochs": 15,
135    "n_train": 400,
136    "n_test": 400,
137    "paired_seeds": [
138      0,
139      1,
140      2,
141      3,
142      4,
143      5,
144      6,
145      7
146    ],
147    "baseline_and_idea_lr_union": [
148      {
149        "lr": 0.001
150      },
151      {
152        "lr": 0.003
153      },
154      {
155        "lr": 0.01
156      }
157    ],
158    "baseline_selection_seeds": [
159      0,
160      1,
161      2,
162      3
163    ],
164    "parameters_baseline": 65,
165    "parameters_idea": 185
166  }
167}