{
  "status": "preregistered before development scoring; reviewed by parent agent",
  "model": "sb3/ppo-MountainCarContinuous-v0",
  "revision": "2111a66bf69db028cd2db582aa004d58a735effd",
  "diagnosis": "Separate pilot24 seeds: motor gain .8 improved reward .932 but steps rose27.5%; velocity gain1.2 reduced opposing action fraction .299 to .254 and mean steps9.0% with reward delta-.015. Joint lower torque and increased velocity sensitivity may improve energy efficiency without excessive slowdown. Input interventions are normalized-space changes and not a proven isolated neural representation. No historical regression or induced fault.",
  "parameterization": {
    "motor_gain": [
      0.7,
      1.0
    ],
    "normalized_velocity_input_gain": [
      1.0,
      1.5
    ]
  },
  "search": {
    "algorithm": "16 scrambled Sobol points scipy.stats.qmc.Sobol seed20261107",
    "candidates": [
      [
        0.8112861780449747,
        1.4064963604323566
      ],
      [
        0.9535653941333293,
        1.0919878273271024
      ],
      [
        0.924029572866857,
        1.3525169072672725
      ],
      [
        0.767652378976345,
        1.147005726583302
      ],
      [
        0.70194558724761,
        1.2928303685039282
      ],
      [
        0.8584956584498287,
        1.2086530178785324
      ],
      [
        0.9630301967263222,
        1.472050140146166
      ],
      [
        0.8205803921446204,
        1.028394800145179
      ],
      [
        0.8464555619284511,
        1.3165697525255382
      ],
      [
        0.9934784077107907,
        1.1819773628376424
      ],
      [
        0.884146367572248,
        1.3872698796913028
      ],
      [
        0.7324714973568915,
        1.1121936412528157
      ],
      [
        0.7417773656547069,
        1.4528272431343794
      ],
      [
        0.8935814825817943,
        1.0485970303416252
      ],
      [
        0.9279145285487175,
        1.2568791531957686
      ],
      [
        0.7807601472362875,
        1.2436287146992981
      ]
    ],
    "budget": 16,
    "selection": "highest dev mean return with zero introduced failure and mean steps <=1.25*original; ties lower index; if none no candidate. No adaptive refinement."
  },
  "splits": {
    "pilot": {
      "start": 0,
      "n": 24
    },
    "development": {
      "start": 10000,
      "n": 128
    },
    "confirmation": {
      "start": 200000,
      "n": 1024
    },
    "parity": {
      "start": 900000,
      "n": 3
    }
  },
  "controls": [
    "original_deterministic",
    "original_stochastic_gSDE_noise_reset_each_episode",
    "fixed_output_gain_0.8_deterministic"
  ],
  "inference": {
    "draws": 50000,
    "seed": 92731,
    "one_sided_alpha": 0.01,
    "return_gate_original": 0.5,
    "return_gate_fixed_output08": 0.1,
    "step_gate": "99% upper bound of candidate steps minus1.25*original steps <=0",
    "failure_gate": "zero introduced failures AND one-sided exact99% Clopper-Pearson upper introduced-failure probability <.005",
    "family": "four gates alpha .01 each; union bound <=.04, bootstrap return/time intervals approximate; binary exact",
    "precision": "pilot paired gain.8 vs original sd1.279; n1024 expected SE.040; zero introduced failures exact99% upper .00449; episodes iid default reset seeds, no robustness guarantee outside reset distribution"
  },
  "normalization": "original frozen mean variance epsilon clip, no reward normalization",
  "deployment": "original/candidate deterministic SB3 predict; candidate wrapper-equivalent input scaling then latent output mean scaling before action clipping",
  "acceptance": "AutoWeight0.5 native gates AND external fixed-baseline/time/exact-binary gates required. If native exports but external fails label provisional/not validated. No second confirmation or candidate retry.",
  "cost_accounting": "all 6*24 pilot,16*128 search,3*128 dev controls, original+selected+fixed08+stochastic1024 confirmation, parity and numerical audit separately disclosed",
  "limitations": "Natural reward efficiency opportunity, not documented regression; mean reward improvement may increase completion time within preregistered25% bound; no claimed benefit of weight edit over mathematically equivalent inference wrapper.",
  "presearch_hashes": {
    "base.npz": "4b42a8e8239a814ad59e7a25924aef38ee282f6481ced60d1d1eebf69ba3df83",
    "normalization.json": "dceab8885c2c8fd0aad3db79fc3d876806b93efc892e0772a1fc6e695634efed",
    "pilot.py": "27aa33f018798f08f837430dc19ab77da905f2a63d07ef65b90777de8c7bf7cb",
    "pilot.json": "d308c97524871a3c39eb3c65eabc6c1da9d521e1cd4f912bfb160ce5c2289003",
    "development.jsonl": "c5034b368aeb0c2952560a8c90ad123b62fca921a169aed9c5b23ae37d7c05a1",
    "confirmation.jsonl": "9b58c6f71c8d40782bb5fef9d0d6cd51614d1b40d5e6322ce0efbfddb08b773a"
  }
}