{
  "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": 0.0011778276093536988
      },
      {
        "cfg": {
          "lr": 0.003
        },
        "mean": 0.0007040829586912878
      },
      {
        "cfg": {
          "lr": 0.01
        },
        "mean": 0.00020319201576057822
      }
    ],
    "full": {
      "mean": 0.00017096688088713563,
      "std": 0.00010126879946081006,
      "per_seed": [
        0.00029771021218039095,
        0.0003086627402808517,
        8.670530223753303e-05,
        0.00011968980834353715,
        6.000406574457884e-05,
        0.00017784061492420733,
        4.760668889502995e-05,
        0.00026951561449095607
      ],
      "n": 8
    },
    "idea_union_sweep_note": "baseline evaluated at every intervention learning rate; final baseline is its tuned best config"
  },
  "idea": {
    "mean": 0.0006723387723468477,
    "std": 0.00030822213357971823,
    "per_seed": [
      0.000751153624150902,
      0.0005644314223900437,
      0.0010974465403705835,
      0.0007808454683981836,
      0.00016198267985600978,
      0.0005583118181675673,
      0.0003655186155810952,
      0.0010990200098603964
    ],
    "n": 8
  },
  "comparison": {
    "delta_mean": 0.0005013718914597121,
    "idea_wins": 0,
    "n_pairs": 8,
    "per_seed_diffs": [
      0.0004534434119705111,
      0.000255768682109192,
      0.0010107412381330505,
      0.0006611556600546464,
      0.00010197861411143094,
      0.0003804712032433599,
      0.00031791192668606527,
      0.0008295043953694403
    ],
    "p_value": 0.0081,
    "mde": 0.0002573449693961294,
    "mde_rel_pct": 150.52328735295612,
    "verdict": "idea worse (significant)",
    "system_worked": false
  },
  "mechanism_signature": {
    "prediction": "A learned mixture should retain separated predictive modes and improve threshold-probability calibration when the trained task is multimodal.",
    "idea_sweep": [
      {
        "cfg": {
          "lr": 0.001
        },
        "mean": 0.0007984692638274282
      },
      {
        "cfg": {
          "lr": 0.003
        },
        "mean": 0.0057232904800912365
      },
      {
        "cfg": {
          "lr": 0.01
        },
        "mean": 0.012347540003247559
      }
    ],
    "idea_nearby_full": {
      "0.001": {
        "mean": 0.0006723387723468477,
        "std": 0.00030822213357971823,
        "per_seed": [
          0.000751153624150902,
          0.0005644314223900437,
          0.0010974465403705835,
          0.0007808454683981836,
          0.00016198267985600978,
          0.0005583118181675673,
          0.0003655186155810952,
          0.0010990200098603964
        ],
        "n": 8
      },
      "0.003": {
        "mean": 0.07285039079579292,
        "std": 0.15517869917738045,
        "per_seed": [
          0.00048457394586876035,
          0.00356022990308702,
          0.0017170965438708663,
          0.0171312615275383,
          0.003046586411073804,
          0.07597032189369202,
          0.4785449206829071,
          0.002348135458305478
        ],
        "n": 8
      },
      "0.01": {
        "mean": 0.07427592801104765,
        "std": 0.17273816889805907,
        "per_seed": [
          0.04037663713097572,
          0.0033035422675311565,
          0.0027636890299618244,
          0.002946291584521532,
          0.0012332660844549537,
          0.007603917736560106,
          0.5301817059516907,
          0.005798374302685261
        ],
        "n": 8
      }
    },
    "trained_model_signature": {
      "n": 300,
      "test_mse": 0.000751153624150902,
      "predicted_mean_mode_separation": 0.5563346147537231,
      "predicted_mean_component_sd": 0.01935468427836895,
      "predicted_bimodal_fraction": 0.9566666483879089,
      "mixture_chance_abs_error": 0.015057117938995335,
      "moment_gaussian_chance_abs_error": 0.014270426034927342,
      "observed_event_rate": 0.47,
      "confirmed": false
    }
  },
  "selection": {
    "idea_best_cfg": {
      "lr": 0.001
    },
    "epochs": 15,
    "n_train": 800,
    "n_test": 300,
    "structural_match": "dynamics: controlled pendulum multi-step target"
  }
}