{
  "bench_version": 1,
  "track": "semantic_pushforward_classification",
  "model": "mlp_tiny",
  "metric_direction": "lower is better",
  "n_seeds": 8,
  "baseline": {
    "best_cfg": {
      "lr": 0.003,
      "epochs": 18,
      "temperature": 1.0,
      "weight_decay": 0.0
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.001,
          "epochs": 18,
          "temperature": 0.8,
          "weight_decay": 0.0
        },
        "mean": 0.28562499955296516
      },
      {
        "cfg": {
          "lr": 0.001,
          "epochs": 18,
          "temperature": 1.0,
          "weight_decay": 0.0
        },
        "mean": 0.28999999538064003
      },
      {
        "cfg": {
          "lr": 0.001,
          "epochs": 18,
          "temperature": 1.2,
          "weight_decay": 0.0
        },
        "mean": 0.2906249985098839
      },
      {
        "cfg": {
          "lr": 0.003,
          "epochs": 18,
          "temperature": 0.8,
          "weight_decay": 0.0
        },
        "mean": 0.2812499962747097
      },
      {
        "cfg": {
          "lr": 0.003,
          "epochs": 18,
          "temperature": 1.0,
          "weight_decay": 0.0
        },
        "mean": 0.27812499180436134
      },
      {
        "cfg": {
          "lr": 0.003,
          "epochs": 18,
          "temperature": 1.2,
          "weight_decay": 0.0
        },
        "mean": 0.27812499180436134
      },
      {
        "cfg": {
          "lr": 0.006,
          "epochs": 18,
          "temperature": 0.8,
          "weight_decay": 0.0
        },
        "mean": 0.29624999314546585
      },
      {
        "cfg": {
          "lr": 0.006,
          "epochs": 18,
          "temperature": 1.0,
          "weight_decay": 0.0
        },
        "mean": 0.29749999940395355
      },
      {
        "cfg": {
          "lr": 0.006,
          "epochs": 18,
          "temperature": 1.2,
          "weight_decay": 0.0
        },
        "mean": 0.2862499877810478
      }
    ],
    "full": {
      "mean": 0.27843749336898327,
      "std": 0.044545224984989557,
      "per_seed": [
        0.2824999988079071,
        0.32499998807907104,
        0.32999998331069946,
        0.17499999701976776,
        0.2800000011920929,
        0.2750000059604645,
        0.26749998331069946,
        0.29249998927116394
      ],
      "n": 8
    }
  },
  "idea": {
    "mean": 0.27843749336898327,
    "std": 0.04587580413113304,
    "per_seed": [
      0.2800000011920929,
      0.32249999046325684,
      0.3349999785423279,
      0.17249999940395355,
      0.2775000035762787,
      0.27250000834465027,
      0.26999998092651367,
      0.29749998450279236
    ],
    "n": 8
  },
  "comparison": {
    "delta_mean": 0.0,
    "idea_wins": 5,
    "n_pairs": 8,
    "per_seed_diffs": [
      -0.002499997615814209,
      -0.002499997615814209,
      0.004999995231628418,
      -0.002499997615814209,
      -0.002499997615814209,
      -0.002499997615814209,
      0.002499997615814209,
      0.004999995231628418
    ],
    "p_value": 1.0,
    "mde": 0.002956247180700302,
    "mde_rel_pct": 1.0617274078037706,
    "verdict": "no measurable effect",
    "system_worked": false
  },
  "mechanism_signature": {
    "mechanism_signature": {
      "prediction": "pushforward conserves state mass; calibration learns a nontrivial correction when raw response probabilities are distorted",
      "predicted_vs_observed": {
        "predicted_mass_error": 0.0,
        "observed_mass_error_mean": 1.1920928955078125e-07,
        "observed_abs_calibrated_vs_raw_mean": 0.007499914558138698,
        "learned_a_mean_by_state": [
          1.052509069442749,
          1.069460391998291,
          1.037056803703308
        ],
        "learned_a_std_by_state": [
          0.016113081946969032,
          0.022935794666409492,
          0.013442201539874077
        ]
      },
      "confirmed": true
    },
    "math_check": {
      "aggregation_error_vs_direct": 0.0,
      "mass_conservation_error": 4.440892098500626e-16
    },
    "custom_track": {
      "name": "semantic_pushforward_classification",
      "file": "semantic_track.py",
      "domain": "uncertainty_calibration"
    }
  }
}