{
  "bench_version": 1,
  "track": "dynamics",
  "model": "rnn_small",
  "metric_direction": "lower is better",
  "n_seeds": 8,
  "baseline": {
    "best_cfg": {
      "lr": 0.002,
      "weight_decay": 0.0001
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.0005,
          "weight_decay": 0.0
        },
        "mean": 0.11454633437097073
      },
      {
        "cfg": {
          "lr": 0.0005,
          "weight_decay": 0.0001
        },
        "mean": 0.11465956829488277
      },
      {
        "cfg": {
          "lr": 0.001,
          "weight_decay": 0.0
        },
        "mean": 0.005499854683876038
      },
      {
        "cfg": {
          "lr": 0.001,
          "weight_decay": 0.0001
        },
        "mean": 0.005453515565022826
      },
      {
        "cfg": {
          "lr": 0.002,
          "weight_decay": 0.0
        },
        "mean": 0.004555012041237205
      },
      {
        "cfg": {
          "lr": 0.002,
          "weight_decay": 0.0001
        },
        "mean": 0.004531972575932741
      }
    ],
    "full": {
      "mean": 0.004597172053763643,
      "std": 0.0008625188036902094,
      "per_seed": [
        0.004187586717307568,
        0.006327331997454166,
        0.0038855248130857944,
        0.003727446775883436,
        0.005093815270811319,
        0.005022076889872551,
        0.004930851981043816,
        0.0036027419846504927
      ],
      "n": 8
    },
    "all_union_configs": [
      {
        "lr": 0.0005,
        "weight_decay": 0.0
      },
      {
        "lr": 0.0005,
        "weight_decay": 0.0001
      },
      {
        "lr": 0.001,
        "weight_decay": 0.0
      },
      {
        "lr": 0.001,
        "weight_decay": 0.0001
      },
      {
        "lr": 0.002,
        "weight_decay": 0.0
      },
      {
        "lr": 0.002,
        "weight_decay": 0.0001
      }
    ]
  },
  "idea": {
    "mean": 0.7925681881606579,
    "std": 0.47593743937124783,
    "per_seed": [
      0.9278137683868408,
      1.54276442527771,
      0.2908315360546112,
      1.384081244468689,
      0.39941179752349854,
      0.5426146984100342,
      1.049137830734253,
      0.20389020442962646
    ],
    "n": 8
  },
  "comparison": {
    "delta_mean": 0.7879710161068942,
    "idea_wins": 0,
    "n_pairs": 8,
    "per_seed_diffs": [
      0.9236261816695333,
      1.5364370932802558,
      0.2869460112415254,
      1.3803537976928055,
      0.3943179822526872,
      0.5375926215201616,
      1.0442069787532091,
      0.20028746244497597
    ],
    "p_value": 0.0081,
    "mde": 0.4250928421498136,
    "mde_rel_pct": 9246.833426688821,
    "verdict": "idea worse (significant)",
    "system_worked": false
  },
  "mechanism_signature": {
    "quantity": "sigma(theta) at ||theta||=sqrt(d), measured formula in trained-update state space",
    "predicted_slope": 0.6931471805599453,
    "observed_slope": 0.6931471805599447,
    "max_abs_error": 0.0,
    "confirmed": true,
    "note": "The update mechanism was exercised by trained dynamics-track models; this signature tests coefficient scaling, not sampling ESS."
  },
  "idea_sweep": [
    {
      "cfg": {
        "lr": 0.002,
        "weight_decay": 0.0001,
        "alpha": 0.05
      },
      "full": {
        "mean": 2.3324266374111176,
        "std": 0.4043937352445423,
        "per_seed": [
          2.4503955841064453,
          2.867910385131836,
          2.2843286991119385,
          1.803705096244812,
          2.6125521659851074,
          2.8163137435913086,
          2.085012912750244,
          1.7391945123672485
        ],
        "n": 8
      }
    },
    {
      "cfg": {
        "lr": 0.0005,
        "weight_decay": 0.0001,
        "alpha": 0.05
      },
      "full": {
        "mean": 0.7925681881606579,
        "std": 0.47593743937124783,
        "per_seed": [
          0.9278137683868408,
          1.54276442527771,
          0.2908315360546112,
          1.384081244468689,
          0.39941179752349854,
          0.5426146984100342,
          1.049137830734253,
          0.20389020442962646
        ],
        "n": 8
      }
    },
    {
      "cfg": {
        "lr": 0.002,
        "weight_decay": 0.0001,
        "alpha": 0.05
      },
      "full": {
        "mean": 2.3324266374111176,
        "std": 0.4043937352445423,
        "per_seed": [
          2.4503955841064453,
          2.867910385131836,
          2.2843286991119385,
          1.803705096244812,
          2.6125521659851074,
          2.8163137435913086,
          2.085012912750244,
          1.7391945123672485
        ],
        "n": 8
      }
    }
  ],
  "transfer_note": "No built-in latent/energy sampler track exists; dynamics is the closest structural stability/control track. The intervention is therefore a parameter-space Langevin transfer, not latent-state sampling."
}