{
  "bench_version": 1,
  "track": "dynamics",
  "model": "rnn_small",
  "metric_direction": "lower is better",
  "n_seeds": 8,
  "baseline": {
    "best_cfg": {
      "lr": 0.005,
      "weight_decay": 0.0
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.001,
          "weight_decay": 0.0
        },
        "mean": 0.5572434142231941
      },
      {
        "cfg": {
          "lr": 0.003,
          "weight_decay": 0.0
        },
        "mean": 0.18839699774980545
      },
      {
        "cfg": {
          "lr": 0.005,
          "weight_decay": 0.0
        },
        "mean": 0.04846041998825967
      }
    ],
    "full": {
      "mean": 0.044931408134289086,
      "std": 0.03835479942009323,
      "per_seed": [
        0.11073682457208633,
        0.013142957352101803,
        0.052351340651512146,
        0.01761055737733841,
        0.10532612353563309,
        0.013849301263689995,
        0.029122132807970047,
        0.017312027513980865
      ],
      "n": 8
    }
  },
  "idea": {
    "mean": 0.048401523381471634,
    "std": 0.0418776346225875,
    "per_seed": [
      0.018566235899925232,
      0.14753688871860504,
      0.05397643893957138,
      0.04880054295063019,
      0.061454158276319504,
      0.008690654300153255,
      0.037352874875068665,
      0.010834393091499805
    ],
    "n": 8
  },
  "comparison": {
    "delta_mean": 0.003470115247182548,
    "idea_wins": 4,
    "n_pairs": 8,
    "per_seed_diffs": [
      -0.0921705886721611,
      0.13439393136650324,
      0.0016250982880592346,
      0.03118998557329178,
      -0.04387196525931358,
      -0.005158646963536739,
      0.008230742067098618,
      -0.00647763442248106
    ],
    "p_value": 0.89505,
    "mde": 0.054278585169126606,
    "mde_rel_pct": 120.80321410560086,
    "verdict": "no significant win",
    "system_worked": false
  },
  "mechanism_signature": {
    "predicted_vs_observed": {
      "predicted_aux_response": -0.0,
      "observed_short_fpt_response": -0.25157928466796875,
      "abs_error": 0.25157928466796875,
      "confirmed": false,
      "event_rate": 0.27937498688697815
    },
    "selected_idea_cfg": {
      "lr": 0.005,
      "weight_decay": 0.0
    },
    "regularizer": "soft regenerative FPT over short hidden-state prefixes"
  },
  "idea_grid": [
    {
      "cfg": {
        "lr": 0.001,
        "weight_decay": 0.0
      },
      "full": {
        "mean": 0.607760451734066,
        "std": 0.06258788186798508,
        "per_seed": [
          0.6598727703094482,
          0.7312052845954895,
          0.6191193461418152,
          0.622126042842865,
          0.5374733209609985,
          0.584874153137207,
          0.5832061171531677,
          0.5242065787315369
        ],
        "n": 8
      }
    },
    {
      "cfg": {
        "lr": 0.003,
        "weight_decay": 0.0
      },
      "full": {
        "mean": 0.24948854185640812,
        "std": 0.08010761069435396,
        "per_seed": [
          0.28027984499931335,
          0.4332640767097473,
          0.19582977890968323,
          0.19673879444599152,
          0.1759151667356491,
          0.22718526422977448,
          0.2931903898715973,
          0.19350501894950867
        ],
        "n": 8
      }
    },
    {
      "cfg": {
        "lr": 0.005,
        "weight_decay": 0.0
      },
      "full": {
        "mean": 0.048401523381471634,
        "std": 0.0418776346225875,
        "per_seed": [
          0.018566235899925232,
          0.14753688871860504,
          0.05397643893957138,
          0.04880054295063019,
          0.061454158276319504,
          0.008690654300153255,
          0.037352874875068665,
          0.010834393091499805
        ],
        "n": 8
      }
    }
  ],
  "budget": {
    "epochs": 4,
    "n_train": 200,
    "n_test": 200
  }
}