[antonsoo]

ClassificationKriging.predict returns probability-column indices rather than fitted class labels. Use classes_ to decode the winning column, so nonconsecutive integers and text labels survive prediction and score compares the original labels.

import numpy as np
from sklearn.naive_bayes import GaussianNB
from pykrige.ck import ClassificationKriging

x = np.array([[0., 0.], [0., 1.], [1., 0.], [1., 1.]])
y = np.array([10, 10, 20, 20])
model = ClassificationKriging(
    classification_model=GaussianNB(),
    n_closest_points=2, variogram_parameters=[1., 0.],
)
model.fit(x, x, y[:, None])
print(model.predict(x, x))  # Before: [0 0 1 1]; after: [10 10 20 20]
print(model.score(x, x, y))  # Before: 0.0; after: 1.0

Regressions cover nonzero, negative, text, and zero-based labels with ordinary and universal kriging. On 155 recorded Meuse observations, 18 configurations with original soil IDs or explicitly recoded labels now match independently decoded ILR/kriging predictions; six zero-based controls stay exact. Classifier probabilities and kriged residuals remain unchanged. This corrects label semantics, with no claim of improved statistical modelling.

Validation: the full suite passes (64 tests) with Python 3.12 / NumPy 2.5 / scikit-learn 1.9 and Python 3.10 / NumPy 1.21.6 / scikit-learn 1.0.2. Black, isort, cython-lint, and git diff --check pass.