{
  "bench_version": 1,
  "track": "dynamics",
  "model": "rnn_small",
  "metric_direction": "lower is better",
  "n_seeds": 8,
  "baseline": {
    "best_cfg": {
      "lr": 0.005,
      "weight_decay": 0.0001
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.001,
          "weight_decay": 0.0
        },
        "mean": 0.0014746972592547536
      },
      {
        "cfg": {
          "lr": 0.001,
          "weight_decay": 0.0001
        },
        "mean": 0.0014579330163542181
      },
      {
        "cfg": {
          "lr": 0.003,
          "weight_decay": 0.0
        },
        "mean": 0.0009116442524828017
      },
      {
        "cfg": {
          "lr": 0.003,
          "weight_decay": 0.0001
        },
        "mean": 0.000888297741767019
      },
      {
        "cfg": {
          "lr": 0.005,
          "weight_decay": 0.0
        },
        "mean": 0.000684085811371915
      },
      {
        "cfg": {
          "lr": 0.005,
          "weight_decay": 0.0001
        },
        "mean": 0.0006692118913633749
      }
    ],
    "full": {
      "mean": 0.0006622805813094601,
      "std": 0.00017074866873521193,
      "per_seed": [
        0.0006129607791081071,
        0.0005905718426220119,
        0.0005136265535838902,
        0.0009596883901394904,
        0.0005064630531705916,
        0.0005161517765372992,
        0.0006731264875270426,
        0.0009256557677872479
      ],
      "n": 8
    }
  },
  "idea": {
    "mean": 0.0007774294572300278,
    "std": 0.00018976167730898186,
    "per_seed": [
      0.0007154284394346178,
      0.0007136478670872748,
      0.0007075121975503862,
      0.0011022123508155346,
      0.000592997414059937,
      0.0005738030886277556,
      0.0007325921906158328,
      0.0010812421096488833
    ],
    "n": 8
  },
  "comparison": {
    "delta_mean": 0.00011514887592056766,
    "idea_wins": 0,
    "n_pairs": 8,
    "per_seed_diffs": [
      0.00010246766032651067,
      0.0001230760244652629,
      0.000193885643966496,
      0.00014252396067604423,
      8.653436088934541e-05,
      5.7651312090456486e-05,
      5.946570308879018e-05,
      0.00015558634186163545
    ],
    "p_value": 0.0081,
    "mde": 4.0017114274026e-05,
    "mde_rel_pct": 6.042320340255821,
    "verdict": "idea worse (significant)",
    "system_worked": false
  },
  "mechanism_signature": {
    "prediction": "dissipativity penalty lowers trained local perturbation gain",
    "baseline_mean_local_gain": 0.42203639075160027,
    "idea_mean_local_gain": 0.4056360647082329,
    "relative_reduction": 0.0388599808044046,
    "per_seed": [
      {
        "seed": 0,
        "baseline_gain": 0.40170010924339294,
        "idea_gain": 0.39243876934051514
      },
      {
        "seed": 1,
        "baseline_gain": 0.4355578124523163,
        "idea_gain": 0.40910080075263977
      },
      {
        "seed": 2,
        "baseline_gain": 0.4409540295600891,
        "idea_gain": 0.41530841588974
      },
      {
        "seed": 3,
        "baseline_gain": 0.4010472893714905,
        "idea_gain": 0.3920332193374634
      },
      {
        "seed": 4,
        "baseline_gain": 0.42631199955940247,
        "idea_gain": 0.4098639488220215
      },
      {
        "seed": 5,
        "baseline_gain": 0.4153302311897278,
        "idea_gain": 0.4018923342227936
      },
      {
        "seed": 6,
        "baseline_gain": 0.4280075430870056,
        "idea_gain": 0.4121930003166199
      },
      {
        "seed": 7,
        "baseline_gain": 0.42738211154937744,
        "idea_gain": 0.4122580289840698
      }
    ],
    "confirmed": true
  },
  "idea_sweep": [
    {
      "lr": 0.005,
      "result": {
        "mean": 0.0007774294572300278,
        "std": 0.00018976167730898186,
        "per_seed": [
          0.0007154284394346178,
          0.0007136478670872748,
          0.0007075121975503862,
          0.0011022123508155346,
          0.000592997414059937,
          0.0005738030886277556,
          0.0007325921906158328,
          0.0010812421096488833
        ],
        "n": 8
      }
    },
    {
      "lr": 0.001,
      "result": {
        "mean": 0.0016683561116224155,
        "std": 0.0005233230853334032,
        "per_seed": [
          0.0012873527593910694,
          0.001405279035679996,
          0.0013610285241156816,
          0.001857919036410749,
          0.0012282896786928177,
          0.001367267337627709,
          0.0019549040589481592,
          0.002884808462113142
        ],
        "n": 8
      }
    },
    {
      "lr": 0.003,
      "result": {
        "mean": 0.0009767467927304097,
        "std": 0.00019101000978935268,
        "per_seed": [
          0.0007808486698195338,
          0.0008321516797877848,
          0.0010572962928563356,
          0.0012427946785464883,
          0.0007201741682365537,
          0.0008802887750789523,
          0.0010501790093258023,
          0.0012502410681918263
        ],
        "n": 8
      }
    }
  ],
  "protocol_notes": {
    "matched_structure": "dynamics pendulum rollout; stability/control is the target domain",
    "shared_architecture": "bench rnn_small GRU; only training penalty differs",
    "shared_lr_union": [
      0.001,
      0.003,
      0.005
    ],
    "epochs": 20,
    "batch": 128,
    "penalty": "relu(||d output/d input||_2^2 - target_gain^2)^2",
    "target_gain": 0.35,
    "lambda": 0.15
  }
}