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": 12
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "epochs": 12
 17        },
 18        "mean": 0.16501786187291145
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "epochs": 12
 24        },
 25        "mean": 0.1727909930050373
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.006,
 30          "epochs": 12
 31        },
 32        "mean": 0.17529388144612312
 33      }
 34    ],
 35    "full": {
 36      "mean": 0.16652636602520943,
 37      "std": 0.019344256210385904,
 38      "per_seed": [
 39        0.18950511515140533,
 40        0.15041682124137878,
 41        0.14130990207195282,
 42        0.17883960902690887,
 43        0.15188206732273102,
 44        0.1598561406135559,
 45        0.20002053678035736,
 46        0.16038073599338531
 47      ],
 48      "n": 8
 49    }
 50  },
 51  "idea": {
 52    "mean": 0.23773143999278545,
 53    "std": 0.022669798148183287,
 54    "per_seed": [
 55      0.24208609759807587,
 56      0.21345070004463196,
 57      0.24233895540237427,
 58      0.20291335880756378,
 59      0.27286380529403687,
 60      0.23195308446884155,
 61      0.26794296503067017,
 62      0.22830255329608917
 63    ],
 64    "n": 8,
 65    "best_cfg": {
 66      "lr": 0.001,
 67      "epochs": 12
 68    },
 69    "sweep": [
 70      {
 71        "cfg": {
 72          "lr": 0.001,
 73          "epochs": 12
 74        },
 75        "mean": 0.23773143999278545,
 76        "per_seed": [
 77          0.24208609759807587,
 78          0.21345070004463196,
 79          0.24233895540237427,
 80          0.20291335880756378,
 81          0.27286380529403687,
 82          0.23195308446884155,
 83          0.26794296503067017,
 84          0.22830255329608917
 85        ]
 86      },
 87      {
 88        "cfg": {
 89          "lr": 0.003,
 90          "epochs": 12
 91        },
 92        "mean": 0.25968209840357304,
 93        "per_seed": [
 94          0.28427907824516296,
 95          0.2242421656847,
 96          0.24691039323806763,
 97          0.1911633461713791,
 98          0.319369912147522,
 99          0.28412121534347534,
100          0.29372480511665344,
101          0.23364587128162384
102        ]
103      },
104      {
105        "cfg": {
106          "lr": 0.006,
107          "epochs": 12
108        },
109        "mean": 0.3195491824299097,
110        "per_seed": [
111          0.33699294924736023,
112          0.29003939032554626,
113          0.30690643191337585,
114          0.20158611238002777,
115          0.3861159682273865,
116          0.33679136633872986,
117          0.40094634890556335,
118          0.29701489210128784
119        ]
120      }
121    ]
122  },
123  "comparison": {
124    "delta_mean": 0.07120507396757603,
125    "idea_wins": 0,
126    "n_pairs": 8,
127    "per_seed_diffs": [
128      0.05258098244667053,
129      0.06303387880325317,
130      0.10102905333042145,
131      0.024073749780654907,
132      0.12098173797130585,
133      0.07209694385528564,
134      0.0679224282503128,
135      0.06792181730270386
136    ],
137    "p_value": 0.0081,
138    "mde": 0.02453005366966053,
139    "mde_rel_pct": 14.730432336430782,
140    "verdict": "idea worse (significant)",
141    "system_worked": false
142  },
143  "selection": {
144    "baseline_best_cfg": {
145      "lr": 0.001,
146      "epochs": 12
147    },
148    "idea_best_cfg": {
149      "lr": 0.001,
150      "epochs": 12
151    },
152    "shared_lr_union": [
153      0.001,
154      0.003,
155      0.006
156    ]
157  },
158  "structural_match": "multi-token correlated sequence window",
159  "mechanism_signature": {
160    "quantity": "trained-model pair inclusion mass",
161    "predicted_direction": "compatibility boost increases joint inclusion",
162    "boost": 2.0,
163    "observed_before": 0.039688840508461,
164    "observed_after": 0.22817423939704895,
165    "observed_change": 0.18848538398742676,
166    "confirmed": true
167  },
168  "protocol_note": "Eight paired seeds; three shared learning rates; reduced 1200/400 samples and 12 epochs for runtime."
169}