{
  "bench_version": 1,
  "track": "dynamics",
  "model": "rnn_small",
  "metric_direction": "lower is better",
  "n_seeds": 8,
  "baseline": {
    "best_cfg": {
      "lr": 0.006,
      "score_weight": 0.0
    },
    "sweep": [
      {
        "cfg": {
          "lr": 0.001,
          "score_weight": 0.0
        },
        "mean": 0.004586851398926228
      },
      {
        "cfg": {
          "lr": 0.003,
          "score_weight": 0.0
        },
        "mean": 0.0030439750116784126
      },
      {
        "cfg": {
          "lr": 0.006,
          "score_weight": 0.0
        },
        "mean": 0.0019343624007888138
      }
    ],
    "full": {
      "mean": 0.0018362020200584084,
      "std": 0.0005092446787647265,
      "per_seed": [
        0.0017339912010356784,
        0.0023557941894978285,
        0.0013064503436908126,
        0.002341213868930936,
        0.001163388486020267,
        0.001496884971857071,
        0.0026453696191310883,
        0.0016465234803035855
      ],
      "n": 8
    }
  },
  "idea": {
    "mean": 0.0018397023231955245,
    "std": 0.0005083828899465245,
    "per_seed": [
      0.0017356511671096087,
      0.0023582950234413147,
      0.001307925907894969,
      0.0023415545001626015,
      0.0011693481355905533,
      0.0015009039780125022,
      0.0026489796582609415,
      0.0016549602150917053
    ],
    "n": 8
  },
  "comparison": {
    "delta_mean": 3.500303137116134e-06,
    "idea_wins": 0,
    "n_pairs": 8,
    "per_seed_diffs": [
      1.6599660739302635e-06,
      2.5008339434862137e-06,
      1.4755642041563988e-06,
      3.4063123166561127e-07,
      5.959649570286274e-06,
      4.019006155431271e-06,
      3.6100391298532486e-06,
      8.436734788119793e-06
    ],
    "p_value": 0.0081,
    "mde": 2.2160150645813298e-06,
    "mde_rel_pct": 0.12068470900118276,
    "verdict": "idea worse (significant)",
    "system_worked": false
  },
  "mechanism_signature": {
    "idea_cfg": {
      "lr": 0.006,
      "score_weight": 0.03
    },
    "signature": {
      "prediction": "probability-flow consistency reduces prediction sensitivity to score transport",
      "baseline_abs_sensitivity_mean": 0.0339321275241673,
      "idea_abs_sensitivity_mean": 0.033821433782577515,
      "ratio_idea_over_baseline": 0.9967377895032464,
      "n_models": 8,
      "confirmed": true
    }
  },
  "runtime_seconds": 87.39407682418823,
  "idea_sweep": [
    {
      "cfg": {
        "lr": 0.006,
        "score_weight": 0.03
      },
      "result": {
        "mean": 0.0018397023231955245,
        "std": 0.0005083828899465245,
        "per_seed": [
          0.0017356511671096087,
          0.0023582950234413147,
          0.001307925907894969,
          0.0023415545001626015,
          0.0011693481355905533,
          0.0015009039780125022,
          0.0026489796582609415,
          0.0016549602150917053
        ],
        "n": 8
      }
    },
    {
      "cfg": {
        "lr": 0.006,
        "score_weight": 0.1
      },
      "result": {
        "mean": 0.0018478815909475088,
        "std": 0.0005064124207020823,
        "per_seed": [
          0.0017395142931491137,
          0.002364138374105096,
          0.001311446656472981,
          0.0023424141108989716,
          0.001183195854537189,
          0.001510326866991818,
          0.002657373435795307,
          0.0016746431356295943
        ],
        "n": 8
      }
    },
    {
      "cfg": {
        "lr": 0.001,
        "score_weight": 0.03
      },
      "result": {
        "mean": 0.005814504460431635,
        "std": 0.0022258700624946614,
        "per_seed": [
          0.0073998901061713696,
          0.0020209946669638157,
          0.003580172546207905,
          0.005346550140529871,
          0.005660634022206068,
          0.006507787387818098,
          0.0060682338662445545,
          0.009931772947311401
        ],
        "n": 8
      }
    },
    {
      "cfg": {
        "lr": 0.003,
        "score_weight": 0.03
      },
      "result": {
        "mean": 0.0029023701208643615,
        "std": 0.0012422328212200683,
        "per_seed": [
          0.0010790829546749592,
          0.003418009728193283,
          0.0023018938954919577,
          0.005378886125981808,
          0.002690472174435854,
          0.0023395069874823093,
          0.003974617458879948,
          0.0020364916417747736
        ],
        "n": 8
      }
    }
  ],
  "protocol_notes": "Dynamics chosen because controlled pendulum rollout is explicitly a stability/control task. Both systems use rnn_small, same datasets, epochs and Adam; only the self-supervised score-transport consistency loss differs."
}