Author: Hamza Shah, Independent Researcher, London

Tags: LLM Security, AI Ethics, Alignment

When an assistant is given detailed internal rules, does it genuinely follow policy better, or does it simply learn to sound compliant? This paper examines the gap between behavioural compliance signals and actual policy adherence in large language models under heavily constrained system prompts. Drawing on a series of structured prompt experiments, we distinguish between surface-level compliance theatre — where models produce outputs that appear rule-following — and deeper behavioural policy internalisation. We introduce a taxonomy of four compliance modes and evaluate their observable signatures. Implications for enterprise LLM deployment and AI governance are discussed.

Read the full paper