Every cell in the grid is one decision tree, grown on its own bootstrap resample (sampling with replacement) of a 400-case synthetic training set, considering only a random subset of features at each split — that's what makes it a random forest rather than plain bagging.
- Each split picks the feature/threshold that minimises weighted Gini impurity among the randomly-chosen candidate features.
- Cases left out of a tree's bootstrap sample ("out-of-bag") let that tree vote on data it never trained on — averaged over the whole forest this gives a genuine, un-cheated OOB accuracy estimate, no held-out test set required.
- Cell colour = that tree's live vote on the case you set with the sliders (purple = plaintiff, amber = defendant); brightness = the leaf's confidence. The gauge below is the resulting ensemble probability.