{
  "bench_version": 1,
  "track": "sequence",
  "model": "custom_residual_state",
  "metric_direction": "lower is better",
  "n_seeds": 8,
  "baseline": {
    "best_cfg": {
      "lr": 0.003,
      "clock_scale": 1.0
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.001,
          "clock_scale": 1.0
        },
        "mean": 0.28474123775959015
      },
      {
        "cfg": {
          "lr": 0.003,
          "clock_scale": 1.0
        },
        "mean": 0.18260616809129715
      },
      {
        "cfg": {
          "lr": 0.01,
          "clock_scale": 1.0
        },
        "mean": 0.20043297857046127
      }
    ],
    "full": {
      "mean": 0.17484375648200512,
      "std": 0.025356109025214353,
      "per_seed": [
        0.17150327563285828,
        0.1548469513654709,
        0.1772066205739975,
        0.22686782479286194,
        0.1976654827594757,
        0.13995282351970673,
        0.17249508202075958,
        0.15821199119091034
      ],
      "n": 8
    }
  },
  "idea": {
    "mean": 0.7047909945249557,
    "std": 0.055758244673496424,
    "per_seed": [
      0.7154005765914917,
      0.6783341765403748,
      0.7505328059196472,
      0.6759172081947327,
      0.5878732204437256,
      0.7312020659446716,
      0.78669673204422,
      0.7123711705207825
    ],
    "n": 8,
    "best_cfg": {
      "lr": 0.01,
      "clock_scale": 1.0
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.001,
          "clock_scale": 1.0
        },
        "mean": 0.7530882805585861
      },
      {
        "cfg": {
          "lr": 0.003,
          "clock_scale": 1.0
        },
        "mean": 0.7263866513967514
      },
      {
        "cfg": {
          "lr": 0.01,
          "clock_scale": 1.0
        },
        "mean": 0.7050461918115616
      }
    ]
  },
  "comparison": {
    "delta_mean": 0.5299472380429506,
    "idea_wins": 0,
    "n_pairs": 8,
    "per_seed_diffs": [
      0.5438973009586334,
      0.5234872251749039,
      0.5733261853456497,
      0.4490493834018707,
      0.3902077376842499,
      0.5912492424249649,
      0.6142016500234604,
      0.5541591793298721
    ],
    "p_value": 0.0081,
    "mde": 0.06291635027911027,
    "mde_rel_pct": 35.984327690640534,
    "verdict": "idea worse (significant)",
    "system_worked": false
  },
  "mechanism_signature": {
    "predicted_inactive_state_change": 0.0,
    "observed_baseline_inactive_change": 0.0,
    "observed_idea_inactive_change": 0.0,
    "observed_baseline_event_change": 3.591756582260132,
    "observed_idea_event_change": 1.8746105432510376,
    "idea_active_fraction": 0.9478124976158142,
    "predicted_active_fraction_is_data_dependent": true,
    "confirmed": true
  },
  "protocol_notes": {
    "structural_match": "sequence-level temporal forecasting; recurrent state transition replacement",
    "baseline_method": "explicit residual Euler with one chronological transition per token",
    "idea_method": "implicit residual solve with event-clock deltas",
    "paired_seeds": [
      0,
      1,
      2,
      3,
      4,
      5,
      6,
      7
    ],
    "epochs": 15,
    "n_train": 400,
    "n_test": 200
  }
}