{
  "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": 0.0012489092478062958
      },
      {
        "cfg": {
          "lr": 0.003
        },
        "mean": 0.0007282123697223142
      },
      {
        "cfg": {
          "lr": 0.01
        },
        "mean": 0.0002650022797752172
      }
    ],
    "full": {
      "mean": 0.00020160043823125307,
      "std": 0.00011406837376812385,
      "per_seed": [
        0.0002867693838197738,
        0.00022152646852191538,
        0.00010455308074597269,
        0.00044716018601320684,
        8.912176417652518e-05,
        0.0002234217245131731,
        0.00013277865946292877,
        0.00010747223859652877
      ],
      "n": 8
    }
  },
  "idea": {
    "mean": 0.0008653015684103593,
    "std": 0.00041340883077491884,
    "per_seed": [
      0.0006581585621461272,
      0.0009649579296819866,
      0.0017265599453821778,
      0.0007086805999279022,
      0.000453480111900717,
      0.001295149908401072,
      0.0006053093820810318,
      0.0005101161077618599
    ],
    "n": 8
  },
  "comparison": {
    "delta_mean": 0.0006637011301791063,
    "idea_wins": 0,
    "n_pairs": 8,
    "per_seed_diffs": [
      0.00037138917832635343,
      0.0007434314611600712,
      0.0016220068646362051,
      0.0002615204139146954,
      0.00036435834772419184,
      0.001071728183887899,
      0.00047253072261810303,
      0.0004026438691653311
    ],
    "p_value": 0.0081,
    "mde": 0.0003912591628540673,
    "mde_rel_pct": 194.07654382440347,
    "verdict": "idea worse (significant)",
    "system_worked": false
  },
  "mechanism_signature": {
    "quantity": "finite perturbation amplification of trained memory features",
    "predicted_alpha_bound": 1.0,
    "observed_amplification": 5.219648766797036,
    "n_models": 8,
    "tolerance": "observed <= alpha + 0.15",
    "confirmed": false
  },
  "idea_sweep": [
    {
      "cfg": {
        "lr": 0.01,
        "alpha": 0.8
      },
      "mean": 0.0010894885854213499
    },
    {
      "cfg": {
        "lr": 0.01,
        "alpha": 0.9
      },
      "mean": 0.0009876549775071908
    },
    {
      "cfg": {
        "lr": 0.01,
        "alpha": 1.0
      },
      "mean": 0.0008653015684103593
    }
  ],
  "protocol_notes": {
    "epochs": 15,
    "n_train": 800,
    "n_test": 200,
    "matched_task": "controlled damped pendulum horizon-8",
    "architecture_difference": "GRU recurrence versus two-column A/B memory"
  }
}