{
  "protocol": "Post-run MPS sensitivity audit, separate from the original same-CUDA main comparison",
  "selection": {
    "selected_rendering": "semantic",
    "dev_metrics": {
      "neutral": {
        "accuracy": 0.7714285714285715,
        "macro_f1": 0.7459788544523509,
        "nll": 0.6996133521392657,
        "brier": 0.3261039060421592,
        "ece": 0.12216485276160671
      },
      "semantic": {
        "accuracy": 0.7838095238095237,
        "macro_f1": 0.7599396892684371,
        "nll": 0.718921979422934,
        "brier": 0.3180294403591556,
        "ece": 0.12064957876186524
      }
    },
    "criterion": "highest real-label dev macro accuracy, then lower dev NLL",
    "device": "mps",
    "scope": "Post-run sensitivity audit of Laya choice-key formatting, not a new training experiment. No test-based rendering selection."
  },
  "english_semantic_real_macro": {
    "accuracy": 0.7484148342632099,
    "macro_f1": 0.7329973270680202,
    "nll": 0.6309314045108828,
    "brier": 0.3475500737329656,
    "ece": 0.12349987305363984
  },
  "v2_point_wins_vs_english_semantic": 10,
  "typed_native_teacher_metrics": {
    "accuracy": 0.7475,
    "macro_f1": 0.6203131476624142,
    "nll": 0.8965023905846662,
    "brier": 0.06935947732109757,
    "ece": 0.17598656338286872
  },
  "v2_typed_teacher_metrics": {
    "accuracy": 0.7345,
    "macro_f1": 0.5843823268841445,
    "nll": 0.9071323454613555,
    "brier": 0.07585474596137866,
    "ece": 0.13429560744677668
  },
  "native_typed_accuracy_gap_pp": -1.3000000000000012,
  "caveats": [
    "Development rendering selection used real-label macro accuracy, not test results.",
    "This is a device/rendering sensitivity check; do not merge it into the original same-GPU experiment.",
    "The typed checkpoint has upstream training exposure to the public train split from which these calibration examples were drawn.",
    "Teacher agreement is not real-world correctness; probability matching is described by soft NLL/Brier."
  ]
}