Whitepaper

Dual-Tier Feature Hierarchy in Behavioral Drift Detection

Feature-importance analysis across two training regimes suggests a hierarchy of universal indicators and domain-specific specialists in behavioral drift detection.

5 Mar 2026Manuscriptvalidatedfeaturesdrift-detectionxgboost
CIJ Labs

Abstract

This paper argues that not all drift features play the same role. Some appear to act as universal first responders across domains, while others matter only after the monitoring regime is trained on a specific behavioral domain.

What it establishes

  • Energy Distance emerges as a strong cross-domain indicator.
  • LLM-native training highlights a different set of specialist features.
  • Feature importance changes are structured, not arbitrary.

Why it matters

The paper sharpens how the ABIS feature space should be interpreted. It is not just a large vector. It has hierarchy, and that hierarchy matters when deciding what kind of monitoring claim can transfer across domains.