{
  "status": "PASS",
  "endpoints": 23,
  "splits_per_endpoint": 3,
  "prediction_cells_recomputed": 345,
  "mean_endpoint_mae": {
    "released": 0.374183478126786,
    "control": 0.3702320498982857,
    "a1": 0.3685879578170505,
    "a2": 0.3662318774785543,
    "a3": 0.3588572779249417
  },
  "a3_contrasts": {
    "released": {
      "a3_lower_endpoints": 20,
      "a3_higher_endpoints": 3,
      "macro_mae_reduction_pct": 4.095905110126546
    },
    "control": {
      "a3_lower_endpoints": 17,
      "a3_higher_endpoints": 6,
      "macro_mae_reduction_pct": 3.072335843552435
    },
    "a1": {
      "a3_lower_endpoints": 14,
      "a3_higher_endpoints": 9,
      "macro_mae_reduction_pct": 2.639988552457984
    },
    "a2": {
      "a3_lower_endpoints": 16,
      "a3_higher_endpoints": 7,
      "macro_mae_reduction_pct": 2.013642177842485
    }
  },
  "lowest_mean_endpoints": {
    "released": 1,
    "a1": 7,
    "a2": 3,
    "a3": 12
  },
  "uncertainty": "population SD across the three fixed cluster splits; not SE or CI",
  "limitation": "One pretrained checkpoint per arm and matched training replicate0. Historical released GRASP is a separate reference, not the matched control. Macro MAE mixes endpoint scales."
}