{
  "bench_version": 1,
  "track": "sequence",
  "model": "transformer_tiny",
  "metric_direction": "lower is better",
  "n_seeds": 8,
  "baseline": {
    "best_cfg": {
      "lr": 0.0015,
      "epochs": 10,
      "degree": 8
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.0015,
          "epochs": 10,
          "degree": 8
        },
        "mean": 0.3802499696612358
      },
      {
        "cfg": {
          "lr": 0.003,
          "epochs": 10,
          "degree": 8
        },
        "mean": 0.4271381199359894
      },
      {
        "cfg": {
          "lr": 0.006,
          "epochs": 10,
          "degree": 8
        },
        "mean": 0.5638246461749077
      }
    ],
    "full": {
      "mean": 0.37510566785931587,
      "std": 0.05712290975923622,
      "per_seed": [
        0.45052582025527954,
        0.3252289593219757,
        0.38266339898109436,
        0.36258170008659363,
        0.40209755301475525,
        0.29846498370170593,
        0.46274900436401367,
        0.3165339231491089
      ],
      "n": 8
    }
  },
  "idea": {
    "mean": 0.376313004642725,
    "std": 0.05670116290415251,
    "per_seed": [
      0.46788883209228516,
      0.3152424097061157,
      0.38778379559516907,
      0.3687755763530731,
      0.4005262553691864,
      0.3052693009376526,
      0.4445805251598358,
      0.32043734192848206
    ],
    "n": 8
  },
  "comparison": {
    "delta_mean": 0.0012073367834091187,
    "idea_wins": 3,
    "n_pairs": 8,
    "per_seed_diffs": [
      0.017363011837005615,
      -0.009986549615859985,
      0.005120396614074707,
      0.006193876266479492,
      -0.0015712976455688477,
      0.006804317235946655,
      -0.018168479204177856,
      0.003903418779373169
    ],
    "p_value": 0.77295,
    "mde": 0.009204338351285793,
    "mde_rel_pct": 2.453798793234417,
    "verdict": "no measurable effect",
    "system_worked": false
  },
  "mechanism_signature": {
    "graph": "random_regular_plus_self",
    "degree_nonself": 8,
    "attention_layers": 2,
    "idea_lr_sweep": [
      {
        "cfg": {
          "lr": 0.0015,
          "epochs": 10,
          "degree": 8
        },
        "result": {
          "mean": 0.376313004642725,
          "std": 0.05670116290415251,
          "per_seed": [
            0.46788883209228516,
            0.3152424097061157,
            0.38778379559516907,
            0.3687755763530731,
            0.4005262553691864,
            0.3052693009376526,
            0.4445805251598358,
            0.32043734192848206
          ],
          "n": 8
        }
      },
      {
        "cfg": {
          "lr": 0.003,
          "epochs": 10,
          "degree": 8
        },
        "result": {
          "mean": 0.4302271008491516,
          "std": 0.07651031513028718,
          "per_seed": [
            0.4481615424156189,
            0.48518460988998413,
            0.38120338320732117,
            0.44692790508270264,
            0.4615558385848999,
            0.32852280139923096,
            0.5656229853630066,
            0.3246377408504486
          ],
          "n": 8
        }
      },
      {
        "cfg": {
          "lr": 0.006,
          "epochs": 10,
          "degree": 8
        },
        "result": {
          "mean": 0.5617959015071392,
          "std": 0.21138960001988505,
          "per_seed": [
            0.5091874003410339,
            0.671311616897583,
            0.3863237202167511,
            0.542673647403717,
            0.42686569690704346,
            0.3991197943687439,
            1.0728341341018677,
            0.48605120182037354
          ],
          "n": 8
        }
      }
    ],
    "route_growth": {
      "layers": 2,
      "degree_nonself": 8,
      "predicted_reachable_tokens": 17,
      "observed_gradient_sensitive_tokens": 17,
      "predicted_growth_by_layer": [
        1,
        9,
        17
      ],
      "gradient_threshold": 0.00012559862434864045,
      "measurement": "gradient of trained final-layer token-0 latent wrt input positions",
      "confirmed": true
    }
  },
  "protocol_notes": {
    "dataset_sizes": [
      400,
      400
    ],
    "matched_seeds": [
      0,
      1,
      2,
      3,
      4,
      5,
      6,
      7
    ],
    "structural_match": "sequence forecast requires multi-token correlations; attention is the sole architectural change",
    "baseline_sweep_union_lrs": [
      0.0015,
      0.003,
      0.006
    ],
    "idea_best_cfg": {
      "lr": 0.0015,
      "epochs": 10,
      "degree": 8
    }
  }
}