{
  "bench_version": 1,
  "track": "dynamics",
  "model": "rnn_small",
  "metric_direction": "lower is better",
  "n_seeds": 8,
  "baseline": {
    "best_cfg": {
      "lr": 0.003
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.0015
        },
        "mean": 0.01389938301872462
      },
      {
        "cfg": {
          "lr": 0.003
        },
        "mean": 0.006244247197173536
      },
      {
        "cfg": {
          "lr": 0.006
        },
        "mean": 0.007728174212388694
      }
    ],
    "full": {
      "mean": 0.0067647010437212884,
      "std": 0.0039439644249273965,
      "per_seed": [
        0.00582656916230917,
        0.0031395666301250458,
        0.012371733784675598,
        0.0036391192115843296,
        0.002855701372027397,
        0.00469561293721199,
        0.007864921353757381,
        0.013724383898079395
      ],
      "n": 8
    }
  },
  "idea": {
    "mean": 0.013474244042299688,
    "std": 0.005054578092032932,
    "per_seed": [
      0.007837503217160702,
      0.024782177060842514,
      0.007902047596871853,
      0.0145634850487113,
      0.013514268212020397,
      0.010362541303038597,
      0.015141943469643593,
      0.013689986430108547
    ],
    "n": 8
  },
  "comparison": {
    "delta_mean": 0.006709542998578399,
    "idea_wins": 2,
    "n_pairs": 8,
    "per_seed_diffs": [
      0.002010934054851532,
      0.021642610430717468,
      -0.004469686187803745,
      0.01092436583712697,
      0.010658566839993,
      0.005666928365826607,
      0.007277022115886211,
      -3.4397467970848083e-05
    ],
    "p_value": 0.04685,
    "mde": 0.00670636373721558,
    "mde_rel_pct": 99.13762180872938,
    "verdict": "idea worse (significant)",
    "system_worked": false
  },
  "mechanism_signature": {
    "protocol_note": "Complete 8-seed paired protocol; reduced to 400 train/400 test and 8 epochs because standard 4000/30 run exceeded environment time budget.",
    "prediction": "ISS residual cognitive perturbations remain contractive under the conservative bound",
    "signature": {
      "quantity": "trained perturbation norm ratio over 8 steps",
      "baseline_observed_ratio": 1.3777021169662476,
      "idea_observed_ratio": 1.9764628410339355,
      "idea_conservative_predicted_bound": 0.9299999999999999,
      "confirmed": false,
      "note": "GRU signature uses final hidden perturbation; modular uses cognitive trajectory."
    },
    "idea_sweep": [
      {
        "cfg": {
          "lr": 0.003,
          "gamma": 1.5,
          "dt": 0.1
        },
        "mean": 0.013771303230896592
      },
      {
        "cfg": {
          "lr": 0.003,
          "gamma": 2.0,
          "dt": 0.1
        },
        "mean": 0.01809797785244882
      },
      {
        "cfg": {
          "lr": 0.003,
          "gamma": 2.5,
          "dt": 0.1
        },
        "mean": 0.02604409120976925
      }
    ],
    "parameter_counts": {
      "baseline": 13313,
      "idea": 6689
    }
  }
}