{
  "n": 500,
  "argmax_agreement": 0.992,
  "cuda_same_subset": {
    "accuracy": 0.7910177949703642,
    "macro_f1": 0.7760857891780776,
    "nll": 0.6000106706289864,
    "brier": 0.317876961372947,
    "ece": 0.16696329399611096
  },
  "q8_same_subset": {
    "accuracy": 0.7868855635654054,
    "macro_f1": 0.773615445189482,
    "nll": 0.6015417570028033,
    "brier": 0.318930434743987,
    "ece": 0.16630794459650552
  },
  "accuracy_difference": -0.004132231404958775,
  "nll_difference": 0.0015310863738169367,
  "temperature_folded": true,
  "passed": true,
  "practical_deployment_gate_passed": true,
  "initial_strict_accuracy_target_met_on_subset": false,
  "initial_accuracy_loss_target": 0.003,
  "note": "500-example subset check: 99.2% agreement. Observed task-macro loss is 0.413 pp, so the initial 0.3 pp accuracy goal is not met on this subset. Prefer BF16/MLX when accuracy is the priority; this is not a bound on population degradation."
}