{
  "bench_version": 1,
  "track": "dynamics",
  "model": "rnn_small",
  "metric_direction": "lower is better",
  "n_seeds": 8,
  "baseline": {
    "best_cfg": {
      "lr": 0.002
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.001
        },
        "mean": 0.36154668405652046
      },
      {
        "cfg": {
          "lr": 0.002
        },
        "mean": 0.04037633468396962
      },
      {
        "cfg": {
          "lr": 0.003
        },
        "mean": 0.05718149617314339
      }
    ],
    "full": {
      "mean": 0.04037633468396962,
      "std": 0.019611854198954906,
      "per_seed": [
        0.014195265248417854,
        0.05633978545665741,
        0.01657027006149292,
        0.06460500508546829,
        0.021465152502059937,
        0.04150689020752907,
        0.04312236234545708,
        0.06520594656467438
      ],
      "n": 8
    }
  },
  "idea": {
    "per_seed": [
      0.01551117654889822,
      0.05140165239572525,
      0.008671858347952366,
      0.06345316767692566,
      0.025820285081863403,
      0.03666491433978081,
      0.037566497921943665,
      0.061046112328767776
    ],
    "mean": 0.03751695808023214,
    "std": 0.020258781501441284,
    "config": {
      "lr": 0.002,
      "kappa": 0.05,
      "lambda_star": -0.02,
      "gamma_star": 0.05,
      "epochs": 4
    },
    "sweep": [
      {
        "lr": 0.001,
        "kappa": 0.05,
        "mean": 0.35876577347517014,
        "std": 0.07349462586459757
      },
      {
        "lr": 0.001,
        "kappa": 0.2,
        "mean": 0.3647887259721756,
        "std": 0.07371815195644271
      },
      {
        "lr": 0.001,
        "kappa": 0.8,
        "mean": 0.3905522972345352,
        "std": 0.0744036324602034
      },
      {
        "lr": 0.002,
        "kappa": 0.05,
        "mean": 0.03751695808023214,
        "std": 0.020258781501441284
      },
      {
        "lr": 0.002,
        "kappa": 0.2,
        "mean": 0.041668576071970165,
        "std": 0.02693072627574791
      },
      {
        "lr": 0.002,
        "kappa": 0.8,
        "mean": 0.06623788620345294,
        "std": 0.04677960203658622
      },
      {
        "lr": 0.003,
        "kappa": 0.05,
        "mean": 0.06425974145531654,
        "std": 0.016254177274281544
      },
      {
        "lr": 0.003,
        "kappa": 0.2,
        "mean": 0.06531570013612509,
        "std": 0.013804677930261229
      },
      {
        "lr": 0.003,
        "kappa": 0.8,
        "mean": 0.060623346362262964,
        "std": 0.014286012866245238
      }
    ]
  },
  "comparison": {
    "delta_mean": -0.0028593766037374735,
    "idea_wins": 6,
    "n_pairs": 8,
    "per_seed_diffs": [
      0.0013159113004803658,
      -0.004938133060932159,
      -0.007898411713540554,
      -0.001151837408542633,
      0.004355132579803467,
      -0.0048419758677482605,
      -0.0055558644235134125,
      -0.004159834235906601
    ],
    "p_value": 0.10075,
    "mde": 0.0033891705100824727,
    "mde_rel_pct": 8.393952885049904,
    "verdict": "no significant win",
    "system_worked": false
  },
  "mechanism_signature": {
    "track_match": "dynamics stability/control",
    "prediction": {
      "observed_log_angle_slope": -0.035032859627017375,
      "predicted_minus_gap": -0.08474337961524725,
      "observed_gap": 0.08474337961524725,
      "relative_error": 0.5866006313876798,
      "confirmed": false
    },
    "selection": {
      "kappa_candidates": [
        0.05,
        0.2,
        0.8
      ],
      "best_config": {
        "lr": 0.002,
        "kappa": 0.05
      }
    }
  }
}