========================================================================
EXTERNAL BENCHMARK: Folktables ACSIncome (states as natural-shift domains)
========================================================================

Accuracy: in-domain 0.804 [0.785,0.825]  shifted 0.749 [0.739,0.759]

Transfer gap (source-fit ECE - target-fit ECE), per calibrator family:
calibrator              in-domain                shifted
temperature   0.0044 [0.0022,0.0065]   0.0284 [0.0269,0.0299]
platt         0.0018 [-0.0010,0.0047]   0.0478 [0.0436,0.0514]
isotonic      0.0297 [0.0260,0.0337]   0.1079 [0.1014,0.1154]

Paired shifted-minus-in-domain gap increase (temperature, over seeds):
  delta = +0.0241  95% CI [+0.0220, +0.0258]

========================================================================
DOES THE CORE RELATIONSHIP REPLICATE?  (Spearman over shifted pairs)
  confidence-JS  <->  |T_src - T_tgt|  <->  transfer gap
========================================================================
  conf_JS      -> gap            rho=+0.409  (p=2.34e-05)
  |T_src-T_tgt| -> gap           rho=+0.869  (p=1.01e-31)
  conf_JS      -> |T_src-T_tgt|  rho=+0.365  (p=1.89e-04)
  entropy_shift-> gap            rho=+0.284  (p=4.17e-03)
  acc_drop     -> gap            rho=+0.389  (p=6.34e-05)
  ATC_err      -> gap            rho=+0.729  (p=7.57e-18)

========================================================================
FROZEN SENTIMENT WARNING RULE APPLIED OUT-OF-DOMAIN (not tuned here)
  tau_JS = 0.061 (from sentiment);  severe iff gap > 0.03
========================================================================
  shifted pairs=100  severe prevalence=0.25
  TP=0 FP=0 TN=75 FN=25
  raw accuracy=0.750  balanced accuracy=0.500  (sens=0.000, spec=1.000)

  Per-source-state mean placement (mean conf_JS vs frozen tau, mean gap):
    src CA  mean_JS=0.006 (<= tau)  mean_gap=0.0014
    src TX  mean_JS=0.005 (<= tau)  mean_gap=0.0104
    src NY  mean_JS=0.005 (<= tau)  mean_gap=0.0058
    src FL  mean_JS=0.005 (<= tau)  mean_gap=0.0177
    src PR  mean_JS=0.010 (<= tau)  mean_gap=0.1068

  Regime placement summary:
    mean shifted conf-JS = 0.006  (sentiment reference 0.057, MNLI 0.033)
    mean shifted gap     = 0.0284  (sentiment reference 0.026, MNLI 0.014)

========================================================================
REGIME SPLIT: mainland-only vs Puerto-Rico-involving shifts
  (a single benchmark spans both regimes; parameter drift orders them)
========================================================================
  temperature  in-domain=0.0044  mainland=0.0064 [0.0054,0.0076]  PR=0.0615 [0.0581,0.0650]
  platt        in-domain=0.0018  mainland=0.0056 [0.0039,0.0073]  PR=0.1113 [0.1020,0.1186]
  isotonic     in-domain=0.0297  mainland=0.0511 [0.0449,0.0591]  PR=0.1930 [0.1830,0.2027]
  mainland:  |dT|=0.121  acc_drop=0.0049  conf_JS=0.0050  (n=60)
  PR-invol.: |dT|=0.997  acc_drop=0.1166  conf_JS=0.0077  (n=40)
  |T_src-T_tgt| -> gap Spearman rho (all shifted pairs) = 0.869