{
  "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": 0.3383217714726925
      },
      {
        "cfg": {
          "lr": 0.003
        },
        "mean": 0.3591853603720665
      },
      {
        "cfg": {
          "lr": 0.009
        },
        "mean": 1.0258603394031525
      }
    ],
    "full": {
      "mean": 0.3383217714726925,
      "std": 0.046140530839104237,
      "per_seed": [
        0.3308762013912201,
        0.28574880957603455,
        0.3010590970516205,
        0.3668035864830017,
        0.33377403020858765,
        0.28919702768325806,
        0.4317983090877533,
        0.3673171103000641
      ],
      "n": 8
    }
  },
  "idea": {
    "mean": 0.33857686817646027,
    "std": 0.04627190481655083,
    "per_seed": [
      0.3314558267593384,
      0.2857537865638733,
      0.30147865414619446,
      0.36729896068573,
      0.3336082696914673,
      0.28876644372940063,
      0.43178096413612366,
      0.36847203969955444
    ],
    "n": 8
  },
  "comparison": {
    "delta_mean": 0.0002550967037677765,
    "idea_wins": 3,
    "n_pairs": 8,
    "per_seed_diffs": [
      0.0005796253681182861,
      4.976987838745117e-06,
      0.0004195570945739746,
      0.0004953742027282715,
      -0.00016576051712036133,
      -0.0004305839538574219,
      -1.7344951629638672e-05,
      0.0011549293994903564
    ],
    "p_value": 0.2419,
    "mde": 0.00042210225891124443,
    "mde_rel_pct": 0.12476355189140237,
    "verdict": "no measurable effect",
    "system_worked": false
  },
  "idea_sweep": [
    {
      "cfg": {
        "lr": 0.001,
        "lambda_cov": 0.0001
      },
      "result": {
        "mean": 0.33857686817646027,
        "std": 0.04627190481655083,
        "per_seed": [
          0.3314558267593384,
          0.2857537865638733,
          0.30147865414619446,
          0.36729896068573,
          0.3336082696914673,
          0.28876644372940063,
          0.43178096413612366,
          0.36847203969955444
        ],
        "n": 8
      }
    },
    {
      "cfg": {
        "lr": 0.001,
        "lambda_cov": 0.001
      },
      "result": {
        "mean": 0.33870137482881546,
        "std": 0.04630866547362218,
        "per_seed": [
          0.3318597674369812,
          0.28587043285369873,
          0.30147260427474976,
          0.3680233657360077,
          0.3338707685470581,
          0.2883317470550537,
          0.4315110445022583,
          0.3686712682247162
        ],
        "n": 8
      }
    },
    {
      "cfg": {
        "lr": 0.009,
        "lambda_cov": 0.01
      },
      "result": {
        "mean": 1.0103588178753853,
        "std": 0.2665844690585086,
        "per_seed": [
          1.0186030864715576,
          0.9780297875404358,
          0.7393418550491333,
          1.433849811553955,
          1.3804410696029663,
          0.6398636698722839,
          0.8194577097892761,
          1.0732835531234741
        ],
        "n": 8
      }
    }
  ],
  "epochs": 12,
  "batch": 64,
  "mechanism_signature": {
    "quantity": "pooled input-projection covariance deviation on held-out benchmark windows",
    "predicted_finite_width_scale_at_width_64": 0.125,
    "observed_baseline_mean": 0.10427584871649742,
    "observed_idea_mean": 0.10427427012473345,
    "observed_idea_over_baseline": 0.9999848613865684,
    "confirmed": false,
    "note": "The O(n^-1/2) prediction is a scale law, not an absolute covariance value; this signature tests reduction on trained systems."
  }
}