======================================================================
CALIBRATION TRANSFER GAP  (metric: 15-bin ECE unless noted)
======================================================================

Transfer gap per calibrator family (source-fit ECE - target-fit ECE):
calibrator      in-domain gap            shifted gap
temperature   0.0067 [0.0053,0.0078]    0.0283 [0.0218,0.0347]
platt         0.0044 [0.0031,0.0060]    0.0265 [0.0209,0.0323]
isotonic      0.0359 [0.0306,0.0411]    0.0697 [0.0623,0.0746]

Effect size: shifted-minus-indomain transfer gap increase (paired over seed):
  temperature  delta = +0.0216  95% CI [+0.0153, +0.0278]
  platt        delta = +0.0221  95% CI [+0.0151, +0.0286]
  isotonic     delta = +0.0338  95% CI [+0.0212, +0.0434]

======================================================================
METRIC ROBUSTNESS: shifted transfer gap (temperature) across metrics
======================================================================
  ece_10     in=+0.0049  shift=+0.0302  increase=+0.0253 [+0.0197,+0.0309]
  ece_15     in=+0.0067  shift=+0.0283  increase=+0.0216 [+0.0153,+0.0278]
  ece_20     in=+0.0051  shift=+0.0267  increase=+0.0217 [+0.0162,+0.0272]
  ece_30     in=+0.0036  shift=+0.0235  increase=+0.0199 [+0.0149,+0.0241]
  ece_50     in=+0.0041  shift=+0.0196  increase=+0.0154 [+0.0122,+0.0183]
  ece_adapt  in=+0.0030  shift=+0.0207  increase=+0.0177 [+0.0128,+0.0227]
  brier      in=+0.0004  shift=+0.0034  increase=+0.0030 [+0.0022,+0.0039]
  nll        in=+0.0023  shift=+0.0133  increase=+0.0110 [+0.0079,+0.0142]

======================================================================
WHAT PREDICTS THE TRANSFER GAP?  (Spearman rho over shifted pairs)
======================================================================
  gap ~ dT             rho=+0.888  (p=1.22e-41)
  gap ~ conf_js        rho=+0.631  (p=1.14e-14)
  gap ~ entropy_shift  rho=+0.554  (p=5.07e-11)
  gap ~ acc_drop       rho=+0.600  (p=4.31e-13)
  gap ~ atc_err        rho=+0.160  (p=8.09e-02)

ATC label-free accuracy estimate quality (shifted pairs):
  mean |ATC_est - true target acc| = 0.0359  95% CI [0.0310,0.0407]