{
  "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": 0.2947123125195503
      },
      {
        "cfg": {
          "lr": 0.003
        },
        "mean": 0.3950112909078598
      },
      {
        "cfg": {
          "lr": 0.006
        },
        "mean": 0.6112398952245712
      }
    ],
    "full": {
      "mean": 0.3066449798643589,
      "std": 0.032607185412887076,
      "per_seed": [
        0.34020674228668213,
        0.267076700925827,
        0.2600124776363373,
        0.31155332922935486,
        0.3084578514099121,
        0.3220774829387665,
        0.36002540588378906,
        0.28374984860420227
      ],
      "n": 8
    }
  },
  "idea": {
    "mean": 0.30605993419885635,
    "std": 0.03247841729165149,
    "per_seed": [
      0.3398566246032715,
      0.26776447892189026,
      0.2589544951915741,
      0.3121564984321594,
      0.3084918260574341,
      0.3192761540412903,
      0.35937950015068054,
      0.28259989619255066
    ],
    "n": 8
  },
  "comparison": {
    "delta_mean": -0.0005850456655025482,
    "idea_wins": 5,
    "n_pairs": 8,
    "per_seed_diffs": [
      -0.00035011768341064453,
      0.0006877779960632324,
      -0.0010579824447631836,
      0.0006031692028045654,
      3.3974647521972656e-05,
      -0.0028013288974761963,
      -0.0006459057331085205,
      -0.0011499524116516113
    ],
    "p_value": 0.18685,
    "mde": 0.0009443470728781248,
    "mde_rel_pct": 0.3079610412327136,
    "verdict": "no measurable effect",
    "system_worked": false
  },
  "mechanism_signature": {
    "prediction": "dual summary should have low linear overlap with primary summary (abs corr < 0.2), while Gale residual is numerical roundoff",
    "predicted_abs_summary_corr_max": 0.2,
    "observed_abs_summary_corr_mean": 0.1512247179634869,
    "observed_abs_summary_corr_per_seed": [
      0.07177071273326874,
      0.0552356131374836,
      0.29892387986183167,
      0.023424027487635612,
      0.1484474241733551,
      0.4907960593700409,
      0.1091001033782959,
      0.012099923565983772
    ],
    "observed_null_residual_mean": 9.662536939458732e-07,
    "observed_gate_mean": 0.4977433420717716,
    "confirmed": true
  },
  "idea_sweep": [
    {
      "cfg": {
        "lr": 0.001
      },
      "mean": 0.2946830242872238
    },
    {
      "cfg": {
        "lr": 0.003
      },
      "mean": 0.4010865166783333
    },
    {
      "cfg": {
        "lr": 0.006
      },
      "mean": 0.5471484512090683
    }
  ],
  "core_math_check": {
    "sizes": [
      [
        3,
        7
      ],
      [
        8,
        32
      ],
      [
        12,
        32
      ]
    ],
    "residuals": [
      8.81410249758507e-16,
      4.8764372361745095e-15,
      5.559546421959475e-15
    ],
    "max_residual": 5.559546421959475e-15,
    "passed": true
  },
  "setup": {
    "epochs": 8,
    "n_train": 1200,
    "n_test": 400,
    "a": 8,
    "window": 32,
    "track_justification": "sequence is structurally matched because Gale mixing operates across multi-token windows"
  }
}