1{
  2  "bench_version": 1,
  3  "track": "sequence",
  4  "model": "transformer_tiny",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.001
 10    },
 11    "sweep": [
 12      {
 13        "cfg": {
 14          "lr": 0.001
 15        },
 16        "mean": 0.1927255019545555
 17      },
 18      {
 19        "cfg": {
 20          "lr": 0.003
 21        },
 22        "mean": 0.21616622433066368
 23      },
 24      {
 25        "cfg": {
 26          "lr": 0.006
 27        },
 28        "mean": 0.24436583742499352
 29      }
 30    ],
 31    "full": {
 32      "mean": 0.1927262246608734,
 33      "std": 0.022232878429441,
 34      "per_seed": [
 35        0.21683096885681152,
 36        0.1569768339395523,
 37        0.1901477575302124,
 38        0.2069464474916458,
 39        0.19253608584403992,
 40        0.17629599571228027,
 41        0.2282254546880722,
 42        0.17385025322437286
 43      ],
 44      "n": 8
 45    }
 46  },
 47  "idea": {
 48    "mean": 0.46420706808567047,
 49    "std": 0.0696159664906768,
 50    "per_seed": [
 51      0.5837222337722778,
 52      0.46462294459342957,
 53      0.42523160576820374,
 54      0.4688754081726074,
 55      0.47928518056869507,
 56      0.4190153479576111,
 57      0.5342675447463989,
 58      0.33863627910614014
 59    ],
 60    "n": 8
 61  },
 62  "comparison": {
 63    "delta_mean": 0.27148084342479706,
 64    "idea_wins": 0,
 65    "n_pairs": 8,
 66    "per_seed_diffs": [
 67      0.3668912649154663,
 68      0.30764611065387726,
 69      0.23508384823799133,
 70      0.2619289606809616,
 71      0.28674909472465515,
 72      0.2427193522453308,
 73      0.3060420900583267,
 74      0.16478602588176727
 75    ],
 76    "p_value": 0.0081,
 77    "mde": 0.050432977935438064,
 78    "mde_rel_pct": 26.168196893900337,
 79    "verdict": "idea worse (significant)",
 80    "system_worked": false
 81  },
 82  "mechanism_signature": {
 83    "prediction": "softmax floor >= exp(-2 beta)/n and spherical residual norm = 1",
 84    "predicted_floor": 0.004229227601144147,
 85    "observed_floor": 0.015373899950645864,
 86    "max_observed_norm_error": 1.1920928955078125e-07,
 87    "confirmed": true
 88  },
 89  "idea_sweep": [
 90    {
 91      "cfg": {
 92        "lr": 0.001,
 93        "beta": 1.0,
 94        "eta": 0.5
 95      },
 96      "mean": 0.5600788816809654,
 97      "std": 0.06620292359906506
 98    },
 99    {
100      "cfg": {
101        "lr": 0.003,
102        "beta": 1.0,
103        "eta": 0.5
104      },
105      "mean": 0.46420706808567047,
106      "std": 0.0696159664906768
107    },
108    {
109      "cfg": {
110        "lr": 0.006,
111        "beta": 1.0,
112        "eta": 0.5
113      },
114      "mean": 0.4669463559985161,
115      "std": 0.08683987850593247
116    }
117  ],
118  "paired_seed_protocol": {
119    "seeds": [
120      0,
121      1,
122      2,
123      3,
124      4,
125      5,
126      6,
127      7
128    ],
129    "epochs": 8,
130    "n_train": 1200,
131    "n_test": 400
132  },
133  "structural_match": "sequence-level multi-token forecast with attention; same task and end-to-end model family"
134}