{
  "values": {
    "executed_group_size_treatment": 1,
    "executed_group_size_baseline": 8,
    "pilot_mean_reward": 0.5171875,
    "kernel_effective_neighbors_p10": 7.971011470276497,
    "frac_nonzero_advantage_treatment": 1.0,
    "baseline_reward_corr": 0.21788651194150732
  },
  "gate": {
    "passed": false,
    "failed": [
      "kernel_effective_neighbors_p10"
    ],
    "calibration_permitted": true
  },
  "bandwidth": 0.5,
  "note": "Two treatment steps, 128 training rollouts; 512 extra held-out train-prompt samples, no test-set calibration."
}