{
  "bench_version": 1,
  "track": "dynamics",
  "model": "rnn_small",
  "metric_direction": "lower is better",
  "n_seeds": 8,
  "baseline": {
    "best_cfg": {
      "lr": 0.006
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.0015
        },
        "mean": 0.00198724222718738
      },
      {
        "cfg": {
          "lr": 0.003
        },
        "mean": 0.0013494549202732742
      },
      {
        "cfg": {
          "lr": 0.006
        },
        "mean": 0.0007156882056733593
      }
    ],
    "full": {
      "mean": 0.0006826339049439412,
      "std": 0.0002001734108210113,
      "per_seed": [
        0.0008196589769795537,
        0.0005301681812852621,
        0.000536333944182843,
        0.0009765917202457786,
        0.0009811640484258533,
        0.0005039663519710302,
        0.0006512094405479729,
        0.00046197857591323555
      ],
      "n": 8
    }
  },
  "idea": {
    "mean": 0.0005652364307024982,
    "std": 0.00019911234856491642,
    "per_seed": [
      0.0006373166106641293,
      0.0004831327241845429,
      0.0003901786112692207,
      0.0008194606634788215,
      0.0009116221335716546,
      0.0003751792246475816,
      0.0005717286258004606,
      0.00033327285200357437
    ],
    "n": 8
  },
  "comparison": {
    "delta_mean": -0.00011739747424144298,
    "idea_wins": 8,
    "n_pairs": 8,
    "per_seed_diffs": [
      -0.00018234236631542444,
      -4.7035457100719213e-05,
      -0.00014615533291362226,
      -0.00015713105676695704,
      -6.95419148541987e-05,
      -0.00012878712732344866,
      -7.948081474751234e-05,
      -0.00012870572390966117
    ],
    "p_value": 0.0081,
    "mde": 3.942211076832542e-05,
    "mde_rel_pct": 5.775000403995876,
    "verdict": "idea better (significant)",
    "system_worked": true
  },
  "mechanism_signature": {
    "mechanism_signature": {
      "prediction": "encoder-reset should suppress state sensitivity; final perturbation ratio < 1",
      "observed": [
        {
          "mode": "carry",
          "test_mse": 0.001019460498355329,
          "observed_context_to_final_state_ratio": 0.008399932645261288
        },
        {
          "mode": "reset",
          "test_mse": 0.0011074204230681062,
          "observed_context_to_final_state_ratio": 0.008114841766655445
        }
      ],
      "confirmed": true
    },
    "idea_sweep": {
      "0.0015": {
        "mean": 0.001871054424555041,
        "std": 0.0004904266878392466,
        "per_seed": [
          0.0014404993271455169,
          0.001380899571813643,
          0.0018803070997819304,
          0.0025381247978657484,
          0.0022744236048310995,
          0.001058555906638503,
          0.0021054912358522415,
          0.0022901338525116444
        ],
        "n": 8
      },
      "0.003": {
        "mean": 0.0011377038972568698,
        "std": 0.00018237536205938462,
        "per_seed": [
          0.0011074204230681062,
          0.0010416923323646188,
          0.0010215630754828453,
          0.0014193312963470817,
          0.0012926014605909586,
          0.0008008744916878641,
          0.001297146431170404,
          0.00112100166734308
        ],
        "n": 8
      },
      "0.006": {
        "mean": 0.0005652364307024982,
        "std": 0.00019911234856491642,
        "per_seed": [
          0.0006373166106641293,
          0.0004831327241845429,
          0.0003901786112692207,
          0.0008194606634788215,
          0.0009116221335716546,
          0.0003751792246475816,
          0.0005717286258004606,
          0.00033327285200357437
        ],
        "n": 8
      }
    },
    "chosen_lr": 0.006
  },
  "protocol_notes": "8 paired seeds; baseline sweep and idea sweep share lr union; n_train=400, n_test=160, 18 epochs, batch 128."
}