{
  "bench_version": 1,
  "track": "dynamics",
  "model": "rnn_small",
  "metric_direction": "lower is better",
  "n_seeds": 8,
  "baseline": {
    "best_cfg": {
      "lr": 0.01
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.001
        },
        "mean": 0.02141550718806684
      },
      {
        "cfg": {
          "lr": 0.003
        },
        "mean": 0.0037718631356256083
      },
      {
        "cfg": {
          "lr": 0.01
        },
        "mean": 0.0034299326362088323
      }
    ],
    "full": {
      "mean": 0.002775895278318785,
      "std": 0.0017084856346187775,
      "per_seed": [
        0.0014615225372835994,
        0.004334237892180681,
        0.0012712624156847596,
        0.006652707699686289,
        0.0020110509358346462,
        0.0024756770581007004,
        0.0018914247630164027,
        0.0021092789247632027
      ],
      "n": 8
    }
  },
  "idea": {
    "best_config": {
      "lr": 0.01
    },
    "per_seed": [
      0.006583117414265871,
      0.009368831291794777,
      0.007072345819324255,
      0.01220707781612873,
      0.0064491755329072475,
      0.006343720480799675,
      0.006881004199385643,
      0.0053885444067418575
    ],
    "mean": 0.007536727120168507,
    "std": 0.0020598694618333747,
    "n": 8,
    "idea_grid": [
      {
        "lr": 0.001
      },
      {
        "lr": 0.003
      },
      {
        "lr": 0.01
      }
    ]
  },
  "comparison": {
    "delta_mean": 0.004760831841849722,
    "idea_wins": 0,
    "n_pairs": 8,
    "per_seed_diffs": [
      0.005121594876982272,
      0.005034593399614096,
      0.005801083403639495,
      0.005554370116442442,
      0.004438124597072601,
      0.0038680434226989746,
      0.00498957943636924,
      0.003279265481978655
    ],
    "p_value": 0.0081,
    "mde": 0.000711292400072098,
    "mde_rel_pct": 25.623891709016156,
    "verdict": "idea worse (significant)",
    "system_worked": false
  },
  "mechanism_signature": {
    "prediction": "strong contraction and small residual keep observer; otherwise retrieve",
    "observed": {
      "median_r": 0.9880073070526123,
      "mean_switch_rate": 1.0,
      "mean_q_threshold": 0.10375677701085806
    },
    "per_seed": [
      {
        "r_est": 0.9952841997146606,
        "q_threshold": 0.11885927617549896,
        "switch_rate": 1.0,
        "observer_mse": 0.0014615217223763466,
        "retrieval_mse": 0.006583117414265871
      },
      {
        "r_est": 0.980730414390564,
        "q_threshold": 0.13610804080963135,
        "switch_rate": 1.0,
        "observer_mse": 0.004334237426519394,
        "retrieval_mse": 0.009368831291794777
      },
      {
        "r_est": 1.0263248682022095,
        "q_threshold": 0.07391086220741272,
        "switch_rate": 1.0,
        "observer_mse": 0.0012712617171928287,
        "retrieval_mse": 0.007072345819324255
      },
      {
        "r_est": 0.9759218096733093,
        "q_threshold": 0.12216280400753021,
        "switch_rate": 1.0,
        "observer_mse": 0.006652706768363714,
        "retrieval_mse": 0.01220707781612873
      },
      {
        "r_est": 1.007511854171753,
        "q_threshold": 0.10744541883468628,
        "switch_rate": 1.0,
        "observer_mse": 0.0020110502373427153,
        "retrieval_mse": 0.0064491755329072475
      },
      {
        "r_est": 0.9763659238815308,
        "q_threshold": 0.11470293998718262,
        "switch_rate": 1.0,
        "observer_mse": 0.0024756789207458496,
        "retrieval_mse": 0.006343720480799675
      },
      {
        "r_est": 0.9624899625778198,
        "q_threshold": 0.08983396738767624,
        "switch_rate": 1.0,
        "observer_mse": 0.0018914258107542992,
        "retrieval_mse": 0.006881004199385643
      },
      {
        "r_est": 0.9965905547142029,
        "q_threshold": 0.0670309066772461,
        "switch_rate": 1.0,
        "observer_mse": 0.002109278691932559,
        "retrieval_mse": 0.0053885444067418575
      }
    ],
    "confirmed": true
  },
  "protocol_note": "Built-in dynamics is the required stability/control track. Baseline and idea use independently trained rnn_small systems, identical data, epochs, batch, and shared LR grid; gating uses inputs only, never test labels."
}