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": 0.37249068170785904
 17      },
 18      {
 19        "cfg": {
 20          "lr": 0.003
 21        },
 22        "mean": 0.4498218819499016
 23      },
 24      {
 25        "cfg": {
 26          "lr": 0.01
 27        },
 28        "mean": 1.0784044712781906
 29      }
 30    ],
 31    "full": {
 32      "mean": 0.4031560346484184,
 33      "std": 0.052733373521801685,
 34      "per_seed": [
 35        0.3982558250427246,
 36        0.32343214750289917,
 37        0.3480452001094818,
 38        0.42022955417633057,
 39        0.4928964078426361,
 40        0.4031330645084381,
 41        0.4640232026576996,
 42        0.37523287534713745
 43      ],
 44      "n": 8
 45    }
 46  },
 47  "idea": {
 48    "cfg": {
 49      "lr": 0.001
 50    },
 51    "mean": 0.3726405091583729,
 52    "std": 0.06837978903352387,
 53    "per_seed": [
 54      0.3478388786315918,
 55      0.28696209192276,
 56      0.32253625988960266,
 57      0.41124576330184937,
 58      0.4993930757045746,
 59      0.3521648645401001,
 60      0.44697806239128113,
 61      0.3140050768852234
 62    ],
 63    "n": 8
 64  },
 65  "comparison": {
 66    "delta_mean": -0.030515525490045547,
 67    "idea_wins": 7,
 68    "n_pairs": 8,
 69    "per_seed_diffs": [
 70      -0.05041694641113281,
 71      -0.03647005558013916,
 72      -0.02550894021987915,
 73      -0.008983790874481201,
 74      0.0064966678619384766,
 75      -0.05096819996833801,
 76      -0.017045140266418457,
 77      -0.06122779846191406
 78    ],
 79    "p_value": 0.01475,
 80    "mde": 0.019566492866314403,
 81    "mde_rel_pct": 4.853330022302114,
 82    "verdict": "idea better (significant)",
 83    "system_worked": true
 84  },
 85  "mechanism_signature": {
 86    "prediction": "gated local-global attention uses bounded temporal neighbors plus all-token linear statistics",
 87    "observed_mean_gate": 0.5655390620231628,
 88    "observed_local_fraction": 0.4309374988079071,
 89    "predicted_local_edge_degree": 3,
 90    "predicted_global_path": true,
 91    "confirmed": true
 92  },
 93  "idea_sweep": [
 94    {
 95      "cfg": {
 96        "lr": 0.001
 97      },
 98      "mean": 0.3726405091583729,
 99      "std": 0.06837978903352387
100    },
101    {
102      "cfg": {
103        "lr": 0.003
104      },
105      "mean": 0.4466487616300583,
106      "std": 0.09400408679279276
107    },
108    {
109      "cfg": {
110        "lr": 0.01
111      },
112      "mean": 1.0904799774289131,
113      "std": 0.2792352404713769
114    }
115  ],
116  "track_justification": "Sequence forecasting is structurally matched because the intervention replaces multi-token self-attention."
117}