{
  "bench_version": 1,
  "track": "dynamics",
  "model": "rnn_small",
  "metric_direction": "lower is better",
  "n_seeds": 8,
  "baseline": {
    "best_cfg": {
      "lr": 0.006
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.001
        },
        "mean": 0.0013222937704995275
      },
      {
        "cfg": {
          "lr": 0.003
        },
        "mean": 0.0008866888092597947
      },
      {
        "cfg": {
          "lr": 0.006
        },
        "mean": 0.0005522354185814038
      }
    ],
    "full": {
      "mean": 0.000564589565328788,
      "std": 0.00015977200389635955,
      "per_seed": [
        0.00038707852945663035,
        0.0005981649155728519,
        0.00037658659857697785,
        0.0008471116307191551,
        0.0004592438053805381,
        0.0004694555827882141,
        0.0007397530134767294,
        0.0006393224466592073
      ],
      "n": 8
    }
  },
  "idea": {
    "mean": 0.000554656424355926,
    "std": 0.0001615859243021339,
    "per_seed": [
      0.000384580809623003,
      0.0005927970050834119,
      0.000364386651199311,
      0.0008439605007879436,
      0.0004653841897379607,
      0.00047854892909526825,
      0.0007660035626031458,
      0.0005415897467173636
    ],
    "n": 8
  },
  "comparison": {
    "delta_mean": -9.93314097286202e-06,
    "idea_wins": 5,
    "n_pairs": 8,
    "per_seed_diffs": [
      -2.497719833627343e-06,
      -5.367910489439964e-06,
      -1.2199947377666831e-05,
      -3.1511299312114716e-06,
      6.14038435742259e-06,
      9.093346307054162e-06,
      2.6250549126416445e-05,
      -9.773269994184375e-05
    ],
    "p_value": 0.68235,
    "mde": 3.123112811032005e-05,
    "mde_rel_pct": 5.531651668435042,
    "verdict": "no significant win",
    "system_worked": false
  },
  "mechanism_signature": {
    "observed_normalized_variance": 1.0,
    "predicted_normalized_variance": 1.0,
    "observed_weighted_feasible_mean": 2.821424961090088,
    "observed_weighted_violating_mean": -0.26333290338516235,
    "observed_margin": 3.0847578644752502,
    "predicted_margin": 3.0847578644752502,
    "confirmed": true,
    "note": "Signature measured from trained rnn_small behaviour. This is a supervised transfer of trajectory ranking; the canonical dynamics track has no action-policy interface, so exact PPO ratios are not tested."
  },
  "idea_sweep": [
    {
      "cfg": {
        "lr": 0.001,
        "wf": 1.5,
        "wv": 1.0
      },
      "result": {
        "mean": 0.0013214157661423087,
        "std": 0.0003696944861183539,
        "per_seed": [
          0.000921291415579617,
          0.001448000897653401,
          0.001050042686983943,
          0.001866328064352274
        ],
        "n": 4
      }
    },
    {
      "cfg": {
        "lr": 0.003,
        "wf": 2.0,
        "wv": 1.0
      },
      "result": {
        "mean": 0.0008860538364388049,
        "std": 0.0003121798062305589,
        "per_seed": [
          0.0005791240837424994,
          0.0009322315454483032,
          0.0006562279304489493,
          0.0013766317861154675
        ],
        "n": 4
      }
    },
    {
      "cfg": {
        "lr": 0.006,
        "wf": 3.0,
        "wv": 1.0
      },
      "result": {
        "mean": 0.0005464312416734174,
        "std": 0.00019365513117499712,
        "per_seed": [
          0.000384580809623003,
          0.0005927970050834119,
          0.000364386651199311,
          0.0008439605007879436
        ],
        "n": 4
      }
    }
  ],
  "protocol_notes": {
    "selection_seeds": [
      0,
      1,
      2,
      3
    ],
    "paired_seeds": [
      0,
      1,
      2,
      3,
      4,
      5,
      6,
      7
    ],
    "epochs": 10,
    "n_train": 1200,
    "n_test": 400,
    "track_justification": "dynamics is the built-in controlled pendulum multi-step state/action-window task, matching control and terminal feasibility structure."
  }
}