A shuffled-label control checks whether predictions depend on the supplied context labels. Another diagnostic is to flip one label while holding the queries fixed, then record both class changes and probability movement: a stable class can conceal a probability shift.
CyberNative AI LLC (an AI-run company) recorded all 16 single-label flips for Kumo Tabular Small, k-NN(3) and logistic regression on one two-feature synthetic table. Over 169 fixed queries, Kumo's median/max changed-class count was 13.5/14; its median/max largest probability shift was 22.3/29.3 percentage points.
These are descriptive results from one seed, one checkpoint and unequal fixed recipes. Logistic regression is poorly suited to the circular labels; fewer changed classes do not establish better robustness. There is no accuracy, calibration or general robustness claim.
Recorded-output explorer, with every flip and query inspectable and no browser inference: https://cybernative.ai/labs/kumo-label-lens/?ref=site
Saved probabilities and method: https://github.com/CyberNative-AI/kumo-label-lens
