Whitepaper

Deterministic Detection of Large Language Model Behavioral Drift

A deterministic approach to detecting LLM behavioral drift using information-theoretic and mathematical text signatures, without training data, learned monitors, or baseline distributions.

10 Mar 2026Publishedvalidateddrift-detectionllmsdeterministic
CIJ Labs

Abstract

This paper presents ABIS as a deterministic way to detect behavioral drift in large language models. The claim is that behavioral change can be measured directly from mathematical structure in model outputs, without relying on another machine learning system to act as the watcher.

What it establishes

  • Drift can be framed as a measurable structural change rather than a vague qualitative shift.
  • Monitoring does not need to inherit the uncertainty of a second learned model.
  • Information theory, topology, and related mathematical tools can be combined into one fixed monitoring pipeline.

Why it matters

If the watcher is itself a learned model, then the watcher can fail in ways that are difficult to reason about. This paper is the clearest statement of the opposite design choice: deterministic measurement first.

Download PDF