{
  "bench_version": 1,
  "track": "sequence",
  "model": "transformer_tiny",
  "metric_direction": "lower is better",
  "n_seeds": 8,
  "baseline": {
    "best_cfg": {
      "lr": 0.001
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.001
        },
        "mean": 0.16495772451162338
      },
      {
        "cfg": {
          "lr": 0.003
        },
        "mean": 0.18329624831676483
      },
      {
        "cfg": {
          "lr": 0.01
        },
        "mean": 0.43168069422245026
      }
    ],
    "full": {
      "mean": 0.16532287374138832,
      "std": 0.008160238971414253,
      "per_seed": [
        0.17448444664478302,
        0.1530463844537735,
        0.17006324231624603,
        0.16223682463169098,
        0.16158440709114075,
        0.15676827728748322,
        0.17869338393211365,
        0.16570602357387543
      ],
      "n": 8
    }
  },
  "idea": {
    "mean": 0.12441098596900702,
    "std": 0.008560845180504158,
    "per_seed": [
      0.1358136236667633,
      0.10833562165498734,
      0.11411764472723007,
      0.12626340985298157,
      0.1262723207473755,
      0.12286490947008133,
      0.13141369819641113,
      0.1302066594362259
    ],
    "n": 8
  },
  "comparison": {
    "delta_mean": -0.040911887772381306,
    "idea_wins": 8,
    "n_pairs": 8,
    "per_seed_diffs": [
      -0.038670822978019714,
      -0.04471076279878616,
      -0.05594559758901596,
      -0.03597341477870941,
      -0.03531208634376526,
      -0.033903367817401886,
      -0.047279685735702515,
      -0.035499364137649536
    ],
    "p_value": 0.0081,
    "mde": 0.006478177383445342,
    "mde_rel_pct": 3.9185003483420218,
    "verdict": "idea better (significant)",
    "system_worked": true
  },
  "mechanism_signature": {
    "prediction": "hierarchical bounded-neighbor attention preserves measurable long-range influence",
    "baseline_signature": {
      "far_first_token_effect": 0.027809560298919678,
      "local_last_token_effect": 1.0498290061950684,
      "far_to_local_ratio": 0.02648960913626757
    },
    "idea_signature": {
      "far_first_token_effect": 0.020777523517608643,
      "local_last_token_effect": 1.4758778810501099,
      "far_to_local_ratio": 0.014078077626615245
    },
    "confirmed": false,
    "confirmation_rule": "idea far-to-local ratio > baseline far-to-local ratio by at least 20%"
  },
  "extra": {
    "idea_sweep": [
      {
        "cfg": {
          "lr": 0.001
        },
        "result": {
          "mean": 0.12441098596900702,
          "std": 0.008560845180504158,
          "per_seed": [
            0.1358136236667633,
            0.10833562165498734,
            0.11411764472723007,
            0.12626340985298157,
            0.1262723207473755,
            0.12286490947008133,
            0.13141369819641113,
            0.1302066594362259
          ],
          "n": 8
        }
      },
      {
        "cfg": {
          "lr": 0.003
        },
        "result": {
          "mean": 0.13378733582794666,
          "std": 0.012523346697500463,
          "per_seed": [
            0.13517771661281586,
            0.11398444324731827,
            0.1407897174358368,
            0.1486937254667282,
            0.13302475214004517,
            0.11315087229013443,
            0.14389027655124664,
            0.14158718287944794
          ],
          "n": 8
        }
      },
      {
        "cfg": {
          "lr": 0.01
        },
        "result": {
          "mean": 0.1740857046097517,
          "std": 0.015146774808178592,
          "per_seed": [
            0.2027415633201599,
            0.14727982878684998,
            0.17759738862514496,
            0.1726507693529129,
            0.16134709119796753,
            0.1725672036409378,
            0.18487928807735443,
            0.1736225038766861
          ],
          "n": 8
        }
      }
    ],
    "idea_best_cfg": {
      "lr": 0.001
    },
    "params_baseline": 71169,
    "params_idea": 71181,
    "elapsed_sec": 418.1431157588959,
    "n_train": 2000,
    "n_test": 500,
    "epochs": 15
  }
}