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": 0.003,
 10      "epochs": 12
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.0015,
 16          "epochs": 12
 17        },
 18        "mean": 0.3423246815800667
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "epochs": 12
 24        },
 25        "mean": 0.32762672007083893
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.006,
 30          "epochs": 12
 31        },
 32        "mean": 0.5004829466342926
 33      }
 34    ],
 35    "full": {
 36      "mean": 0.32375194132328033,
 37      "std": 0.046624634658664434,
 38      "per_seed": [
 39        0.34710654616355896,
 40        0.2919960618019104,
 41        0.30691948533058167,
 42        0.3644847869873047,
 43        0.31837883591651917,
 44        0.24451658129692078,
 45        0.4092934727668762,
 46        0.3073197603225708
 47      ],
 48      "n": 8
 49    }
 50  },
 51  "idea": {
 52    "mean": 0.3292590510100126,
 53    "std": 0.057546800666266544,
 54    "per_seed": [
 55      0.34751513600349426,
 56      0.24287833273410797,
 57      0.32763731479644775,
 58      0.347289502620697,
 59      0.3484017848968506,
 60      0.23375016450881958,
 61      0.41173285245895386,
 62      0.37486732006073
 63    ],
 64    "n": 8
 65  },
 66  "comparison": {
 67    "delta_mean": 0.005507109686732292,
 68    "idea_wins": 3,
 69    "n_pairs": 8,
 70    "per_seed_diffs": [
 71      0.00040858983993530273,
 72      -0.04911772906780243,
 73      0.02071782946586609,
 74      -0.017195284366607666,
 75      0.03002294898033142,
 76      -0.010766416788101196,
 77      0.0024393796920776367,
 78      0.06754755973815918
 79    ],
 80    "p_value": 0.63405,
 81    "mde": 0.029104389124102455,
 82    "mde_rel_pct": 8.989718796787217,
 83    "verdict": "no significant win",
 84    "system_worked": false
 85  },
 86  "mechanism_signature": {
 87    "idea_sweep": [
 88      {
 89        "cfg": {
 90          "lr": 0.003,
 91          "epochs": 12,
 92          "sigma": 0.35
 93        },
 94        "mean": 0.33411455154418945,
 95        "per_seed": [
 96          0.3964906334877014,
 97          0.2513992488384247,
 98          0.33635154366493225,
 99          0.35221678018569946
100        ]
101      },
102      {
103        "cfg": {
104          "lr": 0.003,
105          "epochs": 12,
106          "sigma": 0.7
107        },
108        "mean": 0.31633007153868675,
109        "per_seed": [
110          0.34751513600349426,
111          0.24287833273410797,
112          0.32763731479644775,
113          0.347289502620697
114        ]
115      },
116      {
117        "cfg": {
118          "lr": 0.003,
119          "epochs": 12,
120          "sigma": 1.4
121        },
122        "mean": 0.3244466111063957,
123        "per_seed": [
124          0.3361342251300812,
125          0.26523905992507935,
126          0.3481976091861725,
127          0.3482155501842499
128        ]
129      }
130    ],
131    "best_idea_cfg": {
132      "lr": 0.003,
133      "epochs": 12,
134      "sigma": 0.7
135    },
136    "mechanism_signature": {
137      "prediction": "Laplacian rows are normalized, nonnegative, and larger sigma gives less selective attention",
138      "trained_model_observed": {
139        "row_sum_error": 2.384185791015625e-07,
140        "min_attention": 1.5814957521342876e-07,
141        "mean_entropy_first_two_seeds": 2.5159993171691895,
142        "learned_sigma_first_two_seeds": [
143          0.7006007432937622,
144          0.7081389427185059
145        ]
146      },
147      "predicted": {
148        "row_sum_error": 0.0,
149        "min_attention": 0.0
150      },
151      "confirmed": true
152    },
153    "operation_accounting": {
154      "dot_qk_channel_multiplications_per_pair": 32,
155      "lap_qk_channel_multiplications_per_pair": 0,
156      "lap_abs_subtracts_per_pair": 32
157    }
158  },
159  "runtime_note": "400 train/200 test, 12 epochs, batch 128, 8 paired seeds"
160}