{
  "bench_version": 1,
  "track": "dynamics",
  "model": "rnn_small",
  "metric_direction": "lower is better",
  "n_seeds": 8,
  "baseline": {
    "best_cfg": {
      "lr": 0.005,
      "epochs": 20,
      "batch": 128
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.001,
          "epochs": 20,
          "batch": 128
        },
        "mean": 0.022121376357972622
      },
      {
        "cfg": {
          "lr": 0.002,
          "epochs": 20,
          "batch": 128
        },
        "mean": 0.016339546535164118
      },
      {
        "cfg": {
          "lr": 0.003,
          "epochs": 20,
          "batch": 128
        },
        "mean": 0.012747961794957519
      },
      {
        "cfg": {
          "lr": 0.005,
          "epochs": 20,
          "batch": 128
        },
        "mean": 0.009133224608376622
      }
    ],
    "full": {
      "mean": 0.008694025862496346,
      "std": 0.0014232314536370147,
      "per_seed": [
        0.007357772905379534,
        0.010028759948909283,
        0.007755488622933626,
        0.011390876956284046,
        0.006608752999454737,
        0.008554458618164062,
        0.008871559053659439,
        0.008984537795186043
      ],
      "n": 8
    }
  },
  "idea": {
    "mean": 0.004228192905429751,
    "std": 0.000991539149802616,
    "per_seed": [
      0.003527658525854349,
      0.0032258331775665283,
      0.0044794389978051186,
      0.00498495576903224,
      0.0034572044387459755,
      0.003025196958333254,
      0.005306838545948267,
      0.005818416830152273
    ],
    "n": 8,
    "config": {
      "lr": 0.005,
      "epochs": 20,
      "batch": 128
    },
    "records": [
      {
        "seed": 0,
        "lr": 0.005,
        "metric": 0.003527658525854349,
        "last_loss": 0.005181310921907425,
        "inverse_error": 1.1013007679139264e-05,
        "condition_mean": 1.57774817943573,
        "observed_latent_mutation_mean": 1.7316803932189941,
        "predicted_vs_observed_probe": {
          "predicted_inverse_error": 0.0,
          "observed_inverse_error": 1.1013007679139264e-05,
          "observed_exchange_mean": 5.360846996307373
        }
      },
      {
        "seed": 1,
        "lr": 0.005,
        "metric": 0.0032258331775665283,
        "last_loss": 0.002949269898235798,
        "inverse_error": 7.703845767537132e-06,
        "condition_mean": 4.959653377532959,
        "observed_latent_mutation_mean": 1.1563115119934082,
        "predicted_vs_observed_probe": {
          "predicted_inverse_error": 0.0,
          "observed_inverse_error": 7.703845767537132e-06,
          "observed_exchange_mean": 2.966520309448242
        }
      },
      {
        "seed": 2,
        "lr": 0.005,
        "metric": 0.0044794389978051186,
        "last_loss": 0.004607265554368496,
        "inverse_error": 1.0601612302707508e-05,
        "condition_mean": 8.03824234008789,
        "observed_latent_mutation_mean": 1.345355749130249,
        "predicted_vs_observed_probe": {
          "predicted_inverse_error": 0.0,
          "observed_inverse_error": 1.0601612302707508e-05,
          "observed_exchange_mean": 4.600798606872559
        }
      },
      {
        "seed": 3,
        "lr": 0.005,
        "metric": 0.00498495576903224,
        "last_loss": 0.0036369254160672428,
        "inverse_error": 7.407511475321371e-06,
        "condition_mean": 4.844733238220215,
        "observed_latent_mutation_mean": 1.294931173324585,
        "predicted_vs_observed_probe": {
          "predicted_inverse_error": 0.0,
          "observed_inverse_error": 7.407511475321371e-06,
          "observed_exchange_mean": 3.5481743812561035
        }
      },
      {
        "seed": 4,
        "lr": 0.005,
        "metric": 0.0034572044387459755,
        "last_loss": 0.0035951392445713283,
        "inverse_error": 8.508128303219564e-06,
        "condition_mean": 5.653021335601807,
        "observed_latent_mutation_mean": 3.2120485305786133,
        "predicted_vs_observed_probe": {
          "predicted_inverse_error": 0.0,
          "observed_inverse_error": 8.508128303219564e-06,
          "observed_exchange_mean": 17.611797332763672
        }
      },
      {
        "seed": 5,
        "lr": 0.005,
        "metric": 0.003025196958333254,
        "last_loss": 0.004070228198543191,
        "inverse_error": 9.308403605245985e-06,
        "condition_mean": 7.194035053253174,
        "observed_latent_mutation_mean": 0.8057935237884521,
        "predicted_vs_observed_probe": {
          "predicted_inverse_error": 0.0,
          "observed_inverse_error": 9.308403605245985e-06,
          "observed_exchange_mean": 1.7656463384628296
        }
      },
      {
        "seed": 6,
        "lr": 0.005,
        "metric": 0.005306838545948267,
        "last_loss": 0.0037229776568710802,
        "inverse_error": 6.436449439206626e-06,
        "condition_mean": 3.7536144256591797,
        "observed_latent_mutation_mean": 1.0492488145828247,
        "predicted_vs_observed_probe": {
          "predicted_inverse_error": 0.0,
          "observed_inverse_error": 6.436449439206626e-06,
          "observed_exchange_mean": 2.438058853149414
        }
      },
      {
        "seed": 7,
        "lr": 0.005,
        "metric": 0.005818416830152273,
        "last_loss": 0.004587573418393731,
        "inverse_error": 1.0803267286974005e-05,
        "condition_mean": 6.939658164978027,
        "observed_latent_mutation_mean": 0.627605140209198,
        "predicted_vs_observed_probe": {
          "predicted_inverse_error": 0.0,
          "observed_inverse_error": 1.0803267286974005e-05,
          "observed_exchange_mean": 1.0146958827972412
        }
      }
    ],
    "sweep": [
      {
        "cfg": {
          "lr": 0.001,
          "epochs": 20,
          "batch": 128
        },
        "mean": 0.02207344933412969
      },
      {
        "cfg": {
          "lr": 0.002,
          "epochs": 20,
          "batch": 128
        },
        "mean": 0.011823555920273066
      },
      {
        "cfg": {
          "lr": 0.003,
          "epochs": 20,
          "batch": 128
        },
        "mean": 0.007646026671864092
      },
      {
        "cfg": {
          "lr": 0.005,
          "epochs": 20,
          "batch": 128
        },
        "mean": 0.004228192905429751
      }
    ]
  },
  "comparison": {
    "delta_mean": -0.0044658329570665956,
    "idea_wins": 8,
    "n_pairs": 8,
    "per_seed_diffs": [
      -0.0038301143795251846,
      -0.006802926771342754,
      -0.0032760496251285076,
      -0.006405921187251806,
      -0.003151548560708761,
      -0.005529261659830809,
      -0.003564720507711172,
      -0.0031661209650337696
    ],
    "p_value": 0.0081,
    "mde": 0.0012800509368697734,
    "mde_rel_pct": 14.723339418526043,
    "verdict": "idea better (significant)",
    "system_worked": true
  },
  "protocol_note": "8 paired seeds; baseline sweep and idea sweep share lr union; official train_model",
  "mechanism_signature": {
    "predicted_inverse_error": 0.0,
    "observed_inverse_error_mean": 8.972778232418932e-06,
    "observed_inverse_error_max": 1.1013007679139264e-05,
    "observed_condition_mean": 5.370088264346123,
    "confirmed": true
  },
  "custom_track": null
}