{
  "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": 0.1927255019545555
      },
      {
        "cfg": {
          "lr": 0.003
        },
        "mean": 0.21616622433066368
      },
      {
        "cfg": {
          "lr": 0.006
        },
        "mean": 0.24436583742499352
      }
    ],
    "full": {
      "mean": 0.1927262246608734,
      "std": 0.022232878429441,
      "per_seed": [
        0.21683096885681152,
        0.1569768339395523,
        0.1901477575302124,
        0.2069464474916458,
        0.19253608584403992,
        0.17629599571228027,
        0.2282254546880722,
        0.17385025322437286
      ],
      "n": 8
    }
  },
  "idea": {
    "mean": 0.46420706808567047,
    "std": 0.0696159664906768,
    "per_seed": [
      0.5837222337722778,
      0.46462294459342957,
      0.42523160576820374,
      0.4688754081726074,
      0.47928518056869507,
      0.4190153479576111,
      0.5342675447463989,
      0.33863627910614014
    ],
    "n": 8
  },
  "comparison": {
    "delta_mean": 0.27148084342479706,
    "idea_wins": 0,
    "n_pairs": 8,
    "per_seed_diffs": [
      0.3668912649154663,
      0.30764611065387726,
      0.23508384823799133,
      0.2619289606809616,
      0.28674909472465515,
      0.2427193522453308,
      0.3060420900583267,
      0.16478602588176727
    ],
    "p_value": 0.0081,
    "mde": 0.050432977935438064,
    "mde_rel_pct": 26.168196893900337,
    "verdict": "idea worse (significant)",
    "system_worked": false
  },
  "mechanism_signature": {
    "prediction": "softmax floor >= exp(-2 beta)/n and spherical residual norm = 1",
    "predicted_floor": 0.004229227601144147,
    "observed_floor": 0.015373899950645864,
    "max_observed_norm_error": 1.1920928955078125e-07,
    "confirmed": true
  },
  "idea_sweep": [
    {
      "cfg": {
        "lr": 0.001,
        "beta": 1.0,
        "eta": 0.5
      },
      "mean": 0.5600788816809654,
      "std": 0.06620292359906506
    },
    {
      "cfg": {
        "lr": 0.003,
        "beta": 1.0,
        "eta": 0.5
      },
      "mean": 0.46420706808567047,
      "std": 0.0696159664906768
    },
    {
      "cfg": {
        "lr": 0.006,
        "beta": 1.0,
        "eta": 0.5
      },
      "mean": 0.4669463559985161,
      "std": 0.08683987850593247
    }
  ],
  "paired_seed_protocol": {
    "seeds": [
      0,
      1,
      2,
      3,
      4,
      5,
      6,
      7
    ],
    "epochs": 8,
    "n_train": 1200,
    "n_test": 400
  },
  "structural_match": "sequence-level multi-token forecast with attention; same task and end-to-end model family"
}