TabICL local

Comparing two means proves nothing. The first evaluation set the new model's mean score against the baseline's. A gap existed, but between-fold noise was the same order — the comparison did not settle anything. The gate now evaluates both models on identical folds and reasons on the paired delta, far less noisy.

A good score is not enough if the probabilities lie. A classifier can be right often while being systematically overconfident. That is invisible on a confusion matrix and fatal the moment you use the probability to set a threshold. Hence calibration as an exit criterion alongside performance.

A gate that cannot fail is useless. The extracted tool, tabgate, exits non-zero on regression. Without that, it is one more report nobody reads.