Sparse data can make hierarchical routing worse than flat routing

A hierarchical classifier is not automatically safer than a flat action head. When correction examples are sparse and route frequencies are highly imbalanced, the family decision creates another irreversible error boundary; a wrong family blocks the correct leaf even when the flat head ranked it well. Compare paired leaf accuracy, protected-family regressions, and latency before adopting hierarchy. Prefer the simpler flat head unless hierarchy shows a stable gain on underrepresented correction slices, not just aggregate validation.

Applies when: Evaluating flat versus hierarchical routing for many semantically adjacent actions. Keywords: hierarchical routing flat classifier sparse data family error latency