{
  "bench_version": 1,
  "track": "dynamics",
  "model": "rnn_small",
  "metric_direction": "lower is better",
  "n_seeds": 8,
  "baseline": {
    "best_cfg": {
      "lr": 0.03
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.001
        },
        "mean": 0.0013811723329126835
      },
      {
        "cfg": {
          "lr": 0.01
        },
        "mean": 0.00036846242801402695
      },
      {
        "cfg": {
          "lr": 0.03
        },
        "mean": 0.0001452554606657941
      }
    ],
    "full": {
      "mean": 0.00016686083745298674,
      "std": 0.0001211998237425883,
      "per_seed": [
        0.000324782362440601,
        5.0498259952291846e-05,
        6.899042637087405e-05,
        0.00013675079389940947,
        0.00021192267013248056,
        0.0003892787790391594,
        9.448564378544688e-05,
        5.817776400363073e-05
      ],
      "n": 8
    }
  },
  "idea": {
    "best_cfg": {
      "lr": 0.03
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.001
        },
        "mean": 0.0015905930049484596,
        "per_seed": [
          0.001139691099524498,
          0.0013835665304213762,
          0.0010758476564660668,
          0.001925584045238793,
          0.0012607797980308533,
          0.001312664826400578,
          0.0020039849914610386,
          0.0026226250920444727
        ]
      },
      {
        "cfg": {
          "lr": 0.01
        },
        "mean": 0.00038149026113387663,
        "per_seed": [
          0.00029065465787425637,
          0.0005194233381189406,
          0.0002394697949057445,
          0.00042430192115716636,
          0.00026041336241178215,
          0.00043081643525511026,
          0.0004192772030364722,
          0.0004675653763115406
        ]
      },
      {
        "cfg": {
          "lr": 0.03
        },
        "mean": 0.00015556820017081918,
        "per_seed": [
          0.0003302668046671897,
          9.448741184314713e-05,
          0.00012044865434290841,
          9.239319479092956e-05,
          0.00011777623876696452,
          0.00015112041728571057,
          0.00016604004485998303,
          0.00017201283480972052
        ]
      }
    ],
    "mean": 0.00015556820017081918,
    "std": 7.185317458066955e-05,
    "per_seed": [
      0.0003302668046671897,
      9.448741184314713e-05,
      0.00012044865434290841,
      9.239319479092956e-05,
      0.00011777623876696452,
      0.00015112041728571057,
      0.00016604004485998303,
      0.00017201283480972052
    ],
    "n": 8
  },
  "comparison": {
    "delta_mean": -1.1292637282167561e-05,
    "idea_wins": 3,
    "n_pairs": 8,
    "per_seed_diffs": [
      5.484442226588726e-06,
      4.398915189085528e-05,
      5.1458227972034365e-05,
      -4.435759910847992e-05,
      -9.414643136551604e-05,
      -0.00023815836175344884,
      7.155440107453614e-05,
      0.00011383507080608979
    ],
    "p_value": 0.82275,
    "mde": 9.440018542032603e-05,
    "mde_rel_pct": 56.57420090973917,
    "verdict": "no significant win",
    "system_worked": false
  },
  "protocol_notes": "Dynamics pendulum is structurally matched to stability/control. Adam is shared; only per-update controller scaling differs.",
  "mechanism_signature": {
    "predicted_cap_mean": 3.5256107662831653,
    "observed_accepted_eta_mean": 2.748427854677808,
    "proposal_cap_fraction": 0.3080357142857143,
    "observed_update_norm_mean": 1.5009487083821522,
    "gradient_ratio_to_next_loss_change_corr": -0.06326550793870624,
    "n_trained_updates": 672,
    "confirmed": true
  },
  "math_sanity": {
    "cap_monotone_in_S": true,
    "formula": "eta_cap=1+sqrt(1+2S)*sqrt(rho*v+eps)"
  }
}