{
  "bench_version": 1,
  "track": "dynamics",
  "model": "rnn_small",
  "metric_direction": "lower is better",
  "n_seeds": 8,
  "baseline": {
    "best_cfg": {
      "lr": 0.006,
      "epochs": 8,
      "eps": 0.1
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.001,
          "epochs": 8,
          "eps": 0.02
        },
        "mean": 0.041175071615725756
      },
      {
        "cfg": {
          "lr": 0.003,
          "epochs": 8,
          "eps": 0.05
        },
        "mean": 0.0074238189263269305
      },
      {
        "cfg": {
          "lr": 0.006,
          "epochs": 8,
          "eps": 0.1
        },
        "mean": 0.005064746190328151
      }
    ],
    "full": {
      "mean": 0.004203042743029073,
      "std": 0.0010498074458499573,
      "per_seed": [
        0.0038119142409414053,
        0.005174982361495495,
        0.005509356502443552,
        0.005762731656432152,
        0.0028818214777857065,
        0.003411747980862856,
        0.003934694454073906,
        0.0031370932701975107
      ],
      "n": 8
    }
  },
  "idea": {
    "mean": 0.002017019782215357,
    "std": 0.0005818964499691051,
    "per_seed": [
      0.002947135828435421,
      0.0018542188918218017,
      0.002415081951767206,
      0.002639889484271407,
      0.0011361651122570038,
      0.0013343744212761521,
      0.0019472046988084912,
      0.0018620878690853715
    ],
    "n": 8
  },
  "comparison": {
    "delta_mean": -0.002186022960813716,
    "idea_wins": 8,
    "n_pairs": 8,
    "per_seed_diffs": [
      -0.0008647784125059843,
      -0.003320763469673693,
      -0.003094274550676346,
      -0.0031228421721607447,
      -0.0017456563655287027,
      -0.0020773735595867038,
      -0.0019874897552654147,
      -0.0012750054011121392
    ],
    "p_value": 0.0081,
    "mde": 0.0007618359394411209,
    "mde_rel_pct": 18.12581946031024,
    "verdict": "idea better (significant)",
    "system_worked": true
  },
  "mechanism_signature": {
    "mechanism_signature": {
      "predicted": {
        "quadratic_contraction": "Schur elimination yields a finite, damped latent contraction"
      },
      "observed_per_seed": [
        {
          "latent_norm": 1.0471861362457275,
          "last_token_norm": 2.2861626148223877,
          "ratio": 0.45805409005283093
        },
        {
          "latent_norm": 1.1378988027572632,
          "last_token_norm": 2.3819708824157715,
          "ratio": 0.47771314551198024
        },
        {
          "latent_norm": 1.0921956300735474,
          "last_token_norm": 2.118751287460327,
          "ratio": 0.5154902496282178
        },
        {
          "latent_norm": 1.2369407415390015,
          "last_token_norm": 2.503779411315918,
          "ratio": 0.49402944043237623
        },
        {
          "latent_norm": 1.1387560367584229,
          "last_token_norm": 2.656961441040039,
          "ratio": 0.4285933620144071
        },
        {
          "latent_norm": 1.123752474784851,
          "last_token_norm": 2.523226022720337,
          "ratio": 0.4453633819030081
        },
        {
          "latent_norm": 1.0531108379364014,
          "last_token_norm": 2.0771141052246094,
          "ratio": 0.5070067336632986
        },
        {
          "latent_norm": 1.042839527130127,
          "last_token_norm": 2.157944917678833,
          "ratio": 0.48325586004825427
        }
      ],
      "confirmed": true
    },
    "idea_sweep": [
      {
        "cfg": {
          "lr": 0.001,
          "epochs": 8,
          "eps": 0.02
        },
        "mean": 0.06400888785719872
      },
      {
        "cfg": {
          "lr": 0.003,
          "epochs": 8,
          "eps": 0.05
        },
        "mean": 0.003893836576025933
      },
      {
        "cfg": {
          "lr": 0.006,
          "epochs": 8,
          "eps": 0.1
        },
        "mean": 0.002464081539073959
      }
    ],
    "structural_match": "controlled pendulum rollout is a temporal dynamics chain"
  }
}