{
  "bench_version": 1,
  "track": "sequence",
  "model": "transformer_tiny",
  "metric_direction": "lower is better",
  "n_seeds": 8,
  "baseline": {
    "best_cfg": {
      "lr": 0.0015,
      "gamma": 1.0
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.0015,
          "gamma": 1.0
        },
        "mean": 0.21315788850188255
      },
      {
        "cfg": {
          "lr": 0.003,
          "gamma": 1.0
        },
        "mean": 0.2156946398317814
      },
      {
        "cfg": {
          "lr": 0.006,
          "gamma": 1.0
        },
        "mean": 0.25452903658151627
      }
    ],
    "full": {
      "mean": 0.2124029453843832,
      "std": 0.029564272137155054,
      "per_seed": [
        0.22719556093215942,
        0.1741580069065094,
        0.20896968245506287,
        0.24230830371379852,
        0.19200055301189423,
        0.1792486160993576,
        0.2660641670227051,
        0.2092786729335785
      ],
      "n": 8
    }
  },
  "idea": {
    "mean": 0.21240293979644775,
    "std": 0.02956426589305894,
    "per_seed": [
      0.22719556093215942,
      0.1741579920053482,
      0.20896968245506287,
      0.24230830371379852,
      0.19200055301189423,
      0.1792486310005188,
      0.2660641372203827,
      0.2092786580324173
    ],
    "n": 8
  },
  "comparison": {
    "delta_mean": -5.587935447692871e-09,
    "idea_wins": 3,
    "n_pairs": 8,
    "per_seed_diffs": [
      0.0,
      -1.4901161193847656e-08,
      0.0,
      0.0,
      0.0,
      1.4901161193847656e-08,
      -2.9802322387695312e-08,
      -1.4901161193847656e-08
    ],
    "p_value": 0.5058,
    "mde": 1.1414612684177589e-08,
    "mde_rel_pct": 5.374036910609076e-06,
    "verdict": "no measurable effect",
    "system_worked": false
  },
  "protocol_notes": "Matched sequence track; shared 2-block d=64 causal transformer, differing only in relative lag bias: unconstrained table versus 8-positive-exponential mixture.",
  "mechanism_signature": {
    "prediction": "trained mixture log-kernel bias decreases with lag; retained content logits can make total attention nonmonotone",
    "observed_bias_max_first_difference": -0.044586181640625,
    "observed_bias_min_second_difference": 0.0005350112915039062,
    "observed_attention_max_first_difference": 0.030105771829364122,
    "n_models": 8,
    "confirmed": true
  },
  "idea_sweep": [
    {
      "cfg": {
        "lr": 0.0015,
        "gamma": 1.0
      },
      "mean": 0.21315788477659225
    },
    {
      "cfg": {
        "lr": 0.003,
        "gamma": 1.0
      },
      "mean": 0.21569465845823288
    },
    {
      "cfg": {
        "lr": 0.006,
        "gamma": 1.0
      },
      "mean": 0.25452912598848343
    }
  ],
  "idea_config": {
    "lr": 0.0015,
    "gamma": 1.0
  },
  "search_space": {
    "baseline_grid": [
      {
        "lr": 0.0015,
        "gamma": 1.0
      },
      {
        "lr": 0.003,
        "gamma": 1.0
      },
      {
        "lr": 0.006,
        "gamma": 1.0
      }
    ],
    "idea_grid": [
      {
        "lr": 0.0015,
        "gamma": 1.0
      },
      {
        "lr": 0.003,
        "gamma": 1.0
      },
      {
        "lr": 0.006,
        "gamma": 1.0
      }
    ]
  }
}