processed/covtoken/s5_conformal_shift.json

{
  "budget": 0.25,
  "alpha": 0.1,
  "nominal_coverage": 0.9,
  "calibrated_on": "LIDC val",
  "guaranteed_coverage": 0.0,
  "tests": {
    "LIDC_test (in-dist)": {
      "n": 500,
      "empirical_coverage": 1.0,
      "mean_Y": 0.8074,
      "coverage_gap_vs_nominal": 0.1,
      "holds": true
    },
    "KiTS (shift)": {
      "n": 500,
      "empirical_coverage": 1.0,
      "mean_Y": 0.7647,
      "coverage_gap_vs_nominal": 0.1,
      "holds": true
    },
    "pancreas (shift)": {
      "n": 351,
      "empirical_coverage": 1.0,
      "mean_Y": 0.8917,
      "coverage_gap_vs_nominal": 0.1,
      "holds": true
    },
    "liver (shift)": {
      "n": 500,
      "empirical_coverage": 1.0,
      "mean_Y": 0.2703,
      "coverage_gap_vs_nominal": 0.1,
      "holds": true
    }
  },
  "in_distribution_holds": true,
  "worst_shift_coverage": 1.0,
  "shift_degrades_guarantee": false,
  "interpretation": "The LIDC-calibrated retention guarantee HOLDS in-distribution; under CT shift the worst empirical coverage is 1.0 (still holds). Per-dataset recalibration (or adaptive conformal) is the fix where it degrades.",
  "elapsed_s": 993.2
}