{
  "bench_version": 1,
  "track": "dynamics",
  "model": "rnn_small",
  "metric_direction": "lower is better",
  "n_seeds": 8,
  "baseline": {
    "best_cfg": {
      "lr": 0.01,
      "delay": 3,
      "epsilon": 0.5
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.001,
          "delay": 0,
          "epsilon": 0.05
        },
        "mean": 0.6047389209270477
      },
      {
        "cfg": {
          "lr": 0.003,
          "delay": 1,
          "epsilon": 0.18
        },
        "mean": 0.37101951241493225
      },
      {
        "cfg": {
          "lr": 0.01,
          "delay": 3,
          "epsilon": 0.5
        },
        "mean": 0.014778749784454703
      }
    ],
    "full": {
      "mean": 0.012637637846637517,
      "std": 0.008144052987350348,
      "per_seed": [
        0.004888118244707584,
        0.009219375438988209,
        0.01350479293614626,
        0.03150271251797676,
        0.006999698933213949,
        0.008468696847558022,
        0.008145930245518684,
        0.01837177760899067
      ],
      "n": 8
    }
  },
  "idea": {
    "mean": 0.17860826198011637,
    "std": 0.049080342215062335,
    "per_seed": [
      0.12278071790933609,
      0.16866335272789001,
      0.180322527885437,
      0.2821338176727295,
      0.13234758377075195,
      0.1678561568260193,
      0.14913593232631683,
      0.22562600672245026
    ],
    "n": 8
  },
  "comparison": {
    "delta_mean": 0.16597062413347885,
    "idea_wins": 0,
    "n_pairs": 8,
    "per_seed_diffs": [
      0.1178925996646285,
      0.1594439772889018,
      0.16681773494929075,
      0.25063110515475273,
      0.125347884837538,
      0.15938745997846127,
      0.14099000208079815,
      0.2072542291134596
    ],
    "p_value": 0.0081,
    "mde": 0.0367800689889995,
    "mde_rel_pct": 291.0359470285464,
    "verdict": "idea worse (significant)",
    "system_worked": false
  },
  "mechanism_signature": {
    "prediction": "triggered updates reduce transmissions while retaining a positive inter-event gap under delayed actuation",
    "trained_model_signature": {
      "trained_event_count": 1,
      "total_steps": 48,
      "communication_fraction": 0.020833333333333332,
      "min_inter_event_gap": null,
      "predicted_gap_proxy": "positive for nonzero threshold",
      "observed_positive_gap": true,
      "mean_drift_norm": 0.24016375679639168,
      "confirmed": true
    },
    "idea_sweep": [
      {
        "cfg": {
          "lr": 0.001,
          "delay": 0,
          "epsilon": 0.05
        },
        "mean": 0.6575141847133636
      },
      {
        "cfg": {
          "lr": 0.003,
          "delay": 1,
          "epsilon": 0.18
        },
        "mean": 0.526682935655117
      },
      {
        "cfg": {
          "lr": 0.01,
          "delay": 3,
          "epsilon": 0.5
        },
        "mean": 0.17860826198011637
      }
    ],
    "parameter_counts": {
      "baseline": 13313,
      "idea": 13313
    },
    "protocol_note": "baseline and idea use identical rnn_small architecture, SGD lr/delay union, epochs, batches, and paired seeds"
  },
  "protocol": {
    "seeds": [
      0,
      1,
      2,
      3,
      4,
      5,
      6,
      7
    ],
    "epochs": 12,
    "n_train": 400,
    "n_test": 200,
    "batch": 128,
    "grid": [
      {
        "lr": 0.001,
        "delay": 0,
        "epsilon": 0.05
      },
      {
        "lr": 0.003,
        "delay": 1,
        "epsilon": 0.18
      },
      {
        "lr": 0.01,
        "delay": 3,
        "epsilon": 0.5
      }
    ]
  }
}