{
  "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": 0.43008940666913986
      },
      {
        "cfg": {
          "lr": 0.003
        },
        "mean": 0.45614606887102127
      },
      {
        "cfg": {
          "lr": 0.005
        },
        "mean": 0.5769104957580566
      }
    ],
    "full": {
      "mean": 0.43690821528434753,
      "std": 0.04200321451882879,
      "per_seed": [
        0.40570154786109924,
        0.35749006271362305,
        0.4590928256511688,
        0.49807319045066833,
        0.43421313166618347,
        0.44106775522232056,
        0.48329558968544006,
        0.41633161902427673
      ],
      "n": 8
    },
    "sweep_union_lrs": [
      0.001,
      0.003,
      0.005
    ]
  },
  "idea": {
    "mean": 0.4732055962085724,
    "std": 0.05560245931804311,
    "per_seed": [
      0.4548509120941162,
      0.4054109454154968,
      0.4097138047218323,
      0.4698072671890259,
      0.5605213046073914,
      0.5118848085403442,
      0.5427255630493164,
      0.4307301640510559
    ],
    "n": 8,
    "best_cfg": {
      "lr": 0.001
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.001
        },
        "mean": 0.4732055962085724,
        "std": 0.05560245931804311,
        "per_seed": [
          0.4548509120941162,
          0.4054109454154968,
          0.4097138047218323,
          0.4698072671890259,
          0.5605213046073914,
          0.5118848085403442,
          0.5427255630493164,
          0.4307301640510559
        ]
      },
      {
        "cfg": {
          "lr": 0.003
        },
        "mean": 0.6192752122879028,
        "std": 0.09782140112106731,
        "per_seed": [
          0.6161494255065918,
          0.5007503628730774,
          0.5499135255813599,
          0.5674563646316528,
          0.7630049586296082,
          0.7055208683013916,
          0.7415022253990173,
          0.5099039673805237
        ]
      },
      {
        "cfg": {
          "lr": 0.005
        },
        "mean": 0.7139862775802612,
        "std": 0.13291049641408512,
        "per_seed": [
          0.7121991515159607,
          0.5803670287132263,
          0.6345881223678589,
          0.5550000071525574,
          0.9426005482673645,
          0.8525124788284302,
          0.8183250427246094,
          0.6162978410720825
        ]
      }
    ]
  },
  "comparison": {
    "delta_mean": 0.036297380924224854,
    "idea_wins": 2,
    "n_pairs": 8,
    "per_seed_diffs": [
      0.04914936423301697,
      0.04792088270187378,
      -0.04937902092933655,
      -0.028265923261642456,
      0.12630817294120789,
      0.07081705331802368,
      0.05942997336387634,
      0.014398545026779175
    ],
    "p_value": 0.1233,
    "mde": 0.04700065274938183,
    "mde_rel_pct": 10.757557561327378,
    "verdict": "no significant win",
    "system_worked": false
  },
  "mechanism_signature": {
    "track_structure": "multi-token sequence forecast",
    "implementation": "learned DCT-basis positive contraction",
    "signature": {
      "predicted": "axis mixing is non-expansive with spectral norm <=1",
      "observed_mean_ratio": 0.7310447409751756,
      "observed_max_ratio": 0.7322232178834247,
      "predicted_max_singular_mean": 0.7532902434468269,
      "confirmed": true
    }
  }
}