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      "epochs": 18
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "epochs": 18
 17        },
 18        "mean": 0.379791796207428
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "epochs": 18
 24        },
 25        "mean": 0.43505316227674484
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.006,
 30          "epochs": 18
 31        },
 32        "mean": 0.5298176184296608
 33      }
 34    ],
 35    "full": {
 36      "mean": 0.3913392685353756,
 37      "std": 0.03427539180539238,
 38      "per_seed": [
 39        0.3928984999656677,
 40        0.31817880272865295,
 41        0.38973456621170044,
 42        0.4183553159236908,
 43        0.3854232728481293,
 44        0.4126569330692291,
 45        0.44099482893943787,
 46        0.3724719285964966
 47      ],
 48      "n": 8
 49    }
 50  },
 51  "idea": {
 52    "mean": 0.22923117503523827,
 53    "std": 0.03083903278155792,
 54    "per_seed": [
 55      0.2540016770362854,
 56      0.19434939324855804,
 57      0.2264380156993866,
 58      0.25375083088874817,
 59      0.20073968172073364,
 60      0.2046380639076233,
 61      0.2879960834980011,
 62      0.21193565428256989
 63    ],
 64    "n": 8
 65  },
 66  "comparison": {
 67    "delta_mean": -0.16210809350013733,
 68    "idea_wins": 8,
 69    "n_pairs": 8,
 70    "per_seed_diffs": [
 71      -0.13889682292938232,
 72      -0.12382940948009491,
 73      -0.16329655051231384,
 74      -0.16460448503494263,
 75      -0.18468359112739563,
 76      -0.20801886916160583,
 77      -0.15299874544143677,
 78      -0.1605362743139267
 79    ],
 80    "p_value": 0.0081,
 81    "mde": 0.021699918431380794,
 82    "mde_rel_pct": 5.545039860833492,
 83    "verdict": "idea better (significant)",
 84    "system_worked": true
 85  },
 86  "math_sanity": [
 87    {
 88      "R": 8,
 89      "ratio": 0.9789600837710886
 90    },
 91    {
 92      "R": 32,
 93      "ratio": 0.9958617558725931
 94    },
 95    {
 96      "R": 128,
 97      "ratio": 0.9992112587895055
 98    },
 99    {
100      "R": 512,
101      "ratio": 0.9998442233803077
102    }
103  ],
104  "idea_sweep": [
105    {
106      "cfg": {
107        "lr": 0.001,
108        "epochs": 18
109      },
110      "mean": 0.2617868036031723
111    },
112    {
113      "cfg": {
114        "lr": 0.003,
115        "epochs": 18
116      },
117      "mean": 0.23213497921824455
118    },
119    {
120      "cfg": {
121        "lr": 0.006,
122        "epochs": 18
123      },
124      "mean": 0.23712797462940216
125    }
126  ],
127  "track_justification": "Sequence forecast has multi-token correlations and is the mandated structural match for attention/SSM ideas.",
128  "mechanism_signature": {
129    "prediction": "dyadic tail should preserve more distant influence than standard local/attention-free mixing",
130    "baseline_observed": {
131      "near_influence": 0.013201632536947727,
132      "far_influence": 0.058944083750247955,
133      "far_near_ratio": 4.464908683135576
134    },
135    "idea_observed": {
136      "near_influence": 0.006817104294896126,
137      "far_influence": 0.08620527386665344,
138      "far_near_ratio": 12.645438609256532
139    },
140    "predicted_far_near_ratio_idea_gt_baseline": true,
141    "confirmed": true
142  }
143}